* Move .gitlab-ci.yml inside the repo
* Remove submodule with gitlab pipeline

Add `maxset` function that returns the most distal set of locations in a location list, similar to existing `minset` function.

Use `minset` and `maxset` consistently in `most_proximal` and `most_distal` locset expressions respectively.

Split Python wrapper changes and tutorial example from the Python documentation branch.

- extend `append` methods for `sample_tree` Python wrapper to accommodate a flatter representation of samples.
- add a tutorial example (will make more sense as part of larger python docs)

Combine badges on readme and move ci git submodule to `ext/`

* Add functions for determining CVs and weights for membrane voltage interpolation and axial current determination.
* Add unit tests for voltage interpolation and axial current interpolation.
* Accommodate mextent changes.

- Add option to compile Arbor for HIP/CUDA backend using Clang. 
- Add new CMake option `ARB_GPU_COMPILE_TYPE` to distinguish between three possible gpu platforms/builds: `cuda-nvcc`; `cuda-clang`; `hip-clang`
- Add gpu wrapper functions in `arbor/backends/gpu/gpu_api.hpp` and `arborenv/gpu_api.hpp` which call HIP/CUDA functions depending on the platform
- Rename functions and files: cuda -> gpu
- Add downgraded warp primitives for HIP
- Implement `uuid` workaround for HIP
- Set correct alignment and warp size for AMD gpus
- Update installation guide and docs.
- Update pip/setuptools to support new gpu targets

Fixes #833

* Remove interleaved matrix solver on GPU.
* Fix out of bound memory bug when initialising matrix for all back ends
* fix singed-vs-unsigned comparison warning 

Fixes #1014 
Fixes #1013

Fix build errors for the GPU tests introduced in #1011.

* Add Dockerfiles for MPI + CUDA test builds on Daint
* Add bors.toml for CI on Daint
* Add the separate ci repo as submodule for convenience
* Add a CI badge for master to the top of the readme

* Extend repertoire of cable cell probes.
* Add the following scalar (i.e. corresponding to just one raw value)
probe address types for cable cells:
   - `cell_probe_membrane_voltage`, replacing `cell_probe_address::membrane_voltage`;
   - `cell_probe_total_ionic_current_density`, replaing `cell_probe_address::membrane_current`;
   - `cell_probe_density_state`;
   - `cell_probe_point_state`;
   - `cell_probe_ion_current_density`;
   - `cell_prone_ion_int_concentration`;
   - `cell_prone_ion_ext_concentration`.
* Extend backend class interface with `mechanism_field_data` method, that returns a raw (device) pointer to the per-instance values of a mechanism state variable.
* Make unit test for probes backend-generic.
* Add unit tests for new probes.
* Modify `fixed_ica_current.mod` unit test NMODL module to be more generic in naming, so it can be better used in probe unit tests.
* Update examples for new cable cell probes.
* Update Python interface for new probe types

Work around issue caused by some versions of pip skipping the `install` stage when no additional command line args are passed via `--install-option`.

Keep settings that can be modified via install option in a singleton that initializes them to their default.

* Replace code that determines density mechanism parameter values across a CV with code that (hopefully) does it correctly. New code has the advantage also of being a bit simpler.
* Add unit test `fvm_layout.density_norm_area_partial` that catches the bug, and which validates the fix.
* Change old 'segment' terminology in `fvm_layout` test comments to 'branch'.

Fixes #1008.

This PR removes the requirement that all cover points are included in a region.
The motivation is to allow more flexible definion of regions, particularly the proximal and distal sets thereof.
The other motivation is that the author finds it much simpler to reason about, however others find the existing approach easier to reason about.

The changes:
 * `mextent` does not always include points on the cover at fork points.
   * it still enforces that its constituent cables have no intersections and are ordered.
 * a `super` region expression adds the cover points to a region.
 * update `most_proximal` to return the `minset` of the proximal points in a region's cables.
 * fix some cut and paste errors in comments and printing of locset expressions.

* add MANIFEST file that describes all paths required in source distribution so that pip can compile Arbor.
* fixes to `setup.py` to help it work across supported Python versions.

Tested with:
* setup tools and pip
* python 3.7 and 3.8
* pip 19 and pip 20
* Linux and OS X

Some small changes to facilitate smooth PyPi integration:
* update the Python installation guide
* update the license file
* add a `setup.cfg`
* add a stub for the Python readme

Pull out a constant computation from the update function to the initialisation step in the HH mechanism. In the busyring benchmark with hh/pas swapped this results in an overall speed-up
of roughly 20%.

* Replace CV cable representation with one which makes it clear to which CV a fork point 'belongs': every fork point belongs to exactly one CV; if a CV representation has cables on two different branches at a fork point, it must have cables on all branches at that fork point, and the fork point is associated with that CV.
* Rework (and simplify) `cv_geometry_from_ends` to suit new representation.
* Add CV child information to `cv_geometry`.
* Add computed pw-constant axial resistivity over cell branches to `fvm_cv_discretization`, so it is available for voltage/current interpolation.
* Extend `location_cv` method: add a new parameter of enum `cv_prefer::type` which instructs `location_cv` on how to resolve locations on CV boundaries.
* Provide pw-constant over cable integration methods for `embed_pwlin`.
* Split out `unique_in_place` to its own header, with unit tests.
* Unit tests to suit.

A support for Gitpod.io, a free automated
dev environment that makes contributing and generally working on GitHub
projects much easier. It allows anyone to start a ready-to-code dev
environment for any branch, issue and pull request with a single click.

* Implement std::variant workalike.
* Add util::variant and accessors util::get, util::get_if, util::visit,
mimicking C++17 std::variant, but with some missing functionality noted
in the `util/variant.hpp` header.
* variant.hpp` available through public header API.

This might be what is breaking zenodo for us.

Fix warning generated when `delay` and `weight` were accidentally swapped in #984 
Match `cell_connection` delay type to `connection` delay type

Fixed `mech_vec` to work with new API and re-enabled the benchmark. Had to disable some of the tests because of long run-times (function responsible for slow-down is `cv_geometry::location_cv`, called from `fvm_layout.cpp:721` and is probably related to unrealistically large cv counts). 

`model-init` run-times have been observed to be too high when running systems with many synapses.
 
The main culprit is the `sort_by` function (fvm_layout:723). When testing on a model with 1024 cells and 10000 synapses, 65% of model-init is spent in this function. 

Comparing two `std::map<std::string, double>` is really expensive. We can replace the string map `synapse_instance.param_value` with a sorted `std::vector<std::pair<unsigned, double>>`, and have an additional `std::map<std::string, unsigned> param_map`  to keep track of all the param_name -> unsigned mappings.

`model-init` run-time is 2.5x faster (tested on my 4 core  i7-7600U laptop):
```
10000 synapses/cell 
--------------------------------------------------------------------
|  num cells       | model-init before  (s) | model-init after (s) |
--------------------------------------------------------------------
|       1024       |          12            |       5              |
|       2048       |          25.5          |       10.8           |
|       4096       |          55.5          |       23.2           |
|       8192       |          121.8         |       49             |
--------------------------------------------------------------------
```

Provide canonical function for mextent that returns a reduced
mcable_list representation which excludes redundant zero-length
cables.

* Provide GPU aware `execution_contexts` to unit tests that run on the GPU.
* Provide multithreaded `execution_contexts` to unit tests that use `thread_private_spike_store`.
* Provide multithreaded `execution_contexts` to unit tests in `test_fvm_lowered.cpp` as a means to test the parallel implementation of `fvm_lowered_cell_impl::initialize`.

Addresses #982

* Change `time_type` to `double` to support time-step accuracy in longer running simulations.
* Change connection delay to `float` from `time_type` in order to avoid unnecessary memory overheads.

* Replace TCLAP with tinyopt, remove TCLAP from `ext/`.
* Replace `sup/tinyopt.hpp` with copy of external tinyopt, update use of `tinyopt` in unit tests and examples to suit.
* Use tinyopt's 'smallopt' interface in `modcc`, `brunel` example.
* Update tinyopt interface in `single` example.
 Fix our attributions and add HBP acknowledgement

Addresses #557 #695 #803 

* Use the task system to parallelize parts of the model initialization phase, namely cell building,`fvm_cv_discretize` and `fvm_build_mechanism_data`. 

Fixes #634 

Use a wrapper `mextent` around an `mcable_list` with a stronger invariant for the representation of thingified regions, with public intersection and union and intersection-testing operations. Creation of an `mextent` requires a morphology, but any further operations on one do not.

* Implement `mextent`, wrapping an `mcable_list` and enforcing an invariant: all cables on the same branch are distinct; and there is a cable in the extent for every branch that intersects with the region on the morphology tree described by the cable list.
* Recast region union and intersection operations on regions in terms of `mextent` objects.
* Use `mextent` objects as the concrete representation for regions in `mprovider`.
* Modify region/locset expression implementations to accommodate new representation.
* Add `region` ctors that wrap an explicit cable list or `mextent`.
* Add a lazy range-based merge view in `util/mergeview.hpp`.

Remove cell_stats from gap junction, dry run and ring examples.

Fixes #962

Add a `setup.py` for installing Arbor directly with pip/setuptools.

Implement a setuptools extension for CMake in `setup.py`. To be honest, I don't understand entirely how it works, but setuptools is obtuse enough that I don't feel at all guilty about this.

Define additional flags for optionally enabling GPUs, MPI, Vectorization and micro-architecture targets, for more adventurous users.

The documentation is updated with a "howto pip" for more casual users who don't want anything to do with CMake.

Fixes #958 .

* Add missing `#include <cstdint>` in `common_types.hpp`.

* Amend `size_t` type check in `mpi.hpp` to support 32-bit builds.

Replace deprecated `__ARM_NEON__` preprocessor define with `__ARM_NEON`, which should be defined by the compiler provided `ARB_ARCH` is set correctly, e.g. with `-DARB_ARCH=armv8-a+simd `

Improve the Python wrapper generation and installation:
  - install a proper module that can be extended with Python code;
  - give the user more control over where to install the module (e.g. as a user package or in a virtualenv).

During building, the following sub-directory is built in the build director (`CMAKE_BINARY_DIR`)
```
└── python
    └── arbor
            ├── __init__.py
            ├── arbor.so
            └── VERSION
```
This path can then be copied VERBATIM to the target installation path. By default this will be in `CMAKE_INSTALL_PREFIX/lib/python%d.%d/site-packages`.
An additional CMake parameter `ARB_PYTHON_PREFIX` can be used to specify an alternative destination for installing the Python module.

The Python part of the wrapper, implemented in `__init__.py` is currently very limited, only providing `__version__` and `__config__` variables.

The installation guide was updated to cover the Python installation.