* Add platform abstraction for loading DLL/SOs.
* Throw more informative exceptions when handling DLL/SOs.
* More informative errors when dealing with dynamically loaded catalogues.
* Translate `arb::file_not_found_error` to `FileNotFoundError` in Python.

- Add `simulation.clear_samplers` to remove spikes and samples from Python.
- Usecase: long-running simulations; where unlimited growth of sampling memory is an issue.

A `--debug` flag is added so that users can inspect the out-of-tree generated C++ code of their catalogue builds.

Adds the `profiler_initialize(ctx)` and `profiler_summary()` functions to the Python module and the `profiling` key to `arbor.config()`. closes #1685.

- `make install` in CI
- make `build-catalogue` play nice with Python 3.6
- execute `build-catalogue`

Fixes #1666

Limit poisson schedule with a stop time using the same semantics as the regular schedule.
closes #1617

More, subtler fallout of the ABI patchset: When mechanisms are added to empty regions, various, layered segfaults
appear.
1. `set_parameter` will receive a `nullptr` upon linear search, throwing `no such parameter`
2. Guarding this for finite `width` uncovers further need in the ABI interface methods
3. This is caused by early exit in `instantiate` on `width == 0`
4. Getting rid of this is better, as we can treat vacuous mechanism painting like all others...
5. ...but causes another segfault, since we default index arrays to `cv_pos.back()` but `cv_pos` is empty.

This patch fixes all of the above and in addition removes mechanisms without support before reifying them
on the FVM cells. Relevant tests are added.

Expose default (empty) catalogue constructor to Python

Fixes 2 bugs where modcc expects a range variable but gets a non-range variable instead.

    PowBinaryExpressions were being handled separately from other BinaryExpressions and as a result, some crucial analysis was being skipped on the analysis of the lhs and rhs arguments. This PR incorporates the analysis and printing of PowBinaryExpression into BinaryExpression.
    Printing code for a masked AssignmentExpression was not handling the case of a non-range variable on the rhs of the assignment correctly. In that case, an explicit cast to a vector is needed, which is added in this PR.

The duplication arose from slightly more complicated composition than usual but could be resolved by proper MRO, it took me a few passes from its original form as well before I figured out the situation is not as complicated as it seems.

`super()` is equal to `super(cls, self)` where `cls` is obtained from the current function's class scope. If arg 2  is an object it will start an MRO search of the object, but skipping classes before the 1st arg type is encountered. In our case with the following base classes: `(_command_template, base_command_class)` this means we can define our base logic in `_command_template` and use `super()` calls there to traverse the MRO in `(base_command_class,)`, so by simply defining `class install_command(_command_template, install)` and `bdist_wheel_command(_command_template, bdist_wheel)` our MRO issue is solved and the duplication removed.

- Allow CV policies to be parsed and written as a `defaultable` in an ACC decor description. 
- The format used is `(default (cv-policy (_cv-policy-sexpr_)))`. 
- Add unit tests. 
- Fix strange white-space formatting.

Fixes #1660

* Filter results of `ls::on_components(1, reg)` to avoid duplicate or
coincident locations arising from terminal zero-length cables in region
argument.

* Fielding an add_sampler call while having no probes is not an error.

Reduce number of parallel build jobs to reduce risk of OOM in CI builds.

Neurolucida ASCII files store samples as `(x y z diameter)`, which were incorrectly being
interpreted as `(x y z radius)`.

Emphasize pip as the preferred method of installation.

Github seems to dislike spaces in Markdown URLs.

- Fixes a set of undefined value warnings.
- merge identical `case`s

- Expose `arb_mechanism_kind` to `mechanism_info`.
- While building mechanism data in `fvm_build_mechanism_data`, check that synapses, painted region dynamics and reversal potential methods have the expected `arb_mechanism_kind`. 
- Add unit test. 

* Categorise ASSIGNED RANGE variables as STATE-ish.
* Also use the opportunity to clean up the infoprinter.
* Accidentaly fix bug in propagating parameter ranges from NMODL ➡️ `mechinfo`

Fixes #1623.

* Clean-up AVX SIMD mask setting.
* Remove redundant global constant in the memory lib.
* Fix 'unused variable' warning in `test_probe.cpp`.

- Add a set of test cases to check the behaviour
- Remove the function `(nil)` from the the DSL
- Remove `nil` and `()` as literals
- Add functions `(region-nil)` and `(locset-nil)`

Fixes an issue where `join` (and likely `intersect`) would not work due to the fact that
`nil` values could not be coerced to `region` or `locset`. This occurred while loading 
a NML files.

While investigating it was found that certain calls, eg `(join () ())` are ambiguous and cannot
be resolved without changes in no proportion to the gain. The cause here is that
we can resolve `()` as region or a locset and at the same time `join` maps a list of
locsets or regions to a locset or region, consequently `(join () ())` can be both.

For future reference, in `label_parse.cpp::eval` we look for the first match in the `eval_map` 
of a function name that can be successfully evaluated. However, this might not be the best match
(and the ordering depends on the internals of `eval_map`).  We _could_ check all successful
evaluations, but as eval is recursive, this idea has some obvious issues.

Fixes hipcc rocm-4.0.1 build for AMD GPU
Fixes #1628

* Run examples via scripts (closes #1566)

- make spack ci run use arbor version information
- Cache Spack for GH Actions