This is a clean-up operation on `fvm_lowered_cell_impl.hpp` and `{cpu,
gpu}/shared_state.{hpp, cpp}`.

- Get rid of `shared_state` specifics in `fvm_lowered_cell_impl` and
push these down to `shared_state`
- Present more of an interface layer towards `fvm_lowered_cell` from
`shared_state`
- De-duplicate methods across CPU/GPU implementations by adding
`shared_state_base`
- Align interfaces of GPU/CPU solvers.

A couple of declarations in `PPACK_IFACE_BLOCK` are fixed for the GPU
backend.

New versions, new warnings. Begone.

- pointers in `PPACK_IFACE_BLOCK` will be `__restrict__`
  - this is expected to produce better binary code for the _scalar_ case
- for GPU we expect this to better utilise data caches as per CUDA docs
- `ved_di` is not used anymore, gone and ABI bumped
- GPU code never used the indexing structs, so remove from
`PPACK_IFACE_BLOCK`

`mc_cell_group::sample_events_` was never used (except clearing it).

Fixes failure to compile with GCC 13.

This is a small PR to track updates in the BluePyOpt API (cf.
https://github.com/BlueBrain/BluePyOpt/pull/393) compared to the v0.8
release.

- Profiler output is now correctly aligned regardless of widths.
- Python (and C++) can now actually set a limit on the nodes printed.

After (with a >=4% filter):
```
REGION                                  CALLS     WALL    THREAD        %
root                                        -    9.401    75.205    100.0
  advance                                   -    9.401    75.205    100.0
    integrate                               -    9.400    75.201    100.0
      current                               -    2.001    16.007     21.3
        zero                         29120728    0.518     4.142      5.5
        Granule_98_3D__all           24000600    0.459     3.675      4.9
        Granule_98_3D__soma_group    24000600    0.456     3.649      4.9
        Gran_CaPool_98               24000600    0.410     3.282      4.4
      state                                 -    1.861    14.888     19.8
        Granule_98_3D__soma_group    24000600    ...

I have been playing around a bit with getting Neural ODE's working (and
in general ANN's as black-box homeostatic mechanisms) and found arbor
lacking in ANN activation functions. Now not everybody might agree that
this is actually needed in Arbor, so I don't expect this PR to be merged
directly. But if people like the idea, here it is :)

The PR contains 3 common unary functions used in ANN's as nonlinear
activations: `sigmoid(x)`, `relu(x)` and `tanh(x)`

Here is a very minimal example of a 'neural ode':

```
NEURON {
  SUFFIX node
}

PARAMETER {
    A1 = -0.662 A2 =  0.282 A3 =  0.957
    A4 = -0.189 A5 = -1.794 A6 =  1.094
    A7 = -1.133 A8 =  0.625 A9 =  0.074
    b1 = -1.434 b2 = -0.358 b3 = -0.093
}

STATE {
    x
    y
    z
}

INITIAL {
    x = 0
    y = 1
    z = 1
}

DERIVATIVE dstate {
    x' = tanh(A1 * x + A2 * y + A3 * z + b1)
    y' = tanh(A4 * x + A5 * y + A6 * z + b2)
 ...

Before
```
✦2 ❯ modcc -t cpu jnml-arbor/SKv3_1.mod
Warnings:
  * jnml-arbor/SKv3_1.mod:(line 1,col 1) Assignments to local variable containing state variables will not be integrated in time
```
After
```
✦2 ❯ modcc -t cpu jnml-arbor/SKv3_1.mod
Warnings:
  * jnml-arbor/SKv3_1.mod:(line 172,col 14) Assignments to local variable containing state variables will not be integrated in time: (rate_m_q = ((m_inf - m_q) / m_tau))
```

Co-authored-by: boeschf <48126478+boeschf@users.noreply.github.com>

`[[maybe_unused]]` in generated code to silence warnings for mechanisms
with `net_receive` methods that do not use the connection weight.

- $x^{-1} \Rightarrow 1/x\quad \forall x$
- $x^n \Rightarrow x\cdot \dots \cdot x \quad x\in N; |x| < 5$
- $x^n \Rightarrow 1/(x\cdot \dots \cdot x) \quad x\in N; |x| < 5; x <
0$
- $b^e \Rightarrow \exp(\log(b) e)\quad \forall  b, e$

The last point introduces potential errors when `pow(b, e)` is allowed,
but `log(b)` is
undefined. These occur exactly when all of the following is true
- $b < 0$
- $e\in N$
- $e$, $b$ not known at compile time (since we cover these cases before)

* Python Wheel workflow uses `ubuntu-20.04` to fix the build
  * In #2058 I forgot to update `ciwheel.yml`
* Why this is necessarily: still have not figured it out. `cibuildwheel`
builds and tests in it's own container...
  * [x] https://github.com/brenthuisman/arbor/actions/runs/3704419712
* Small change to make `musllinux` wheels build. They don't pass testing
yet, so I don't add them to the workflow but that can be solved/added
[later](https://github.com/pypa/cibuildwheel/issues/1376)
* The reason for using `ubuntu-20.04` in the Spack workflow should [no
longer exist](https://github.com/actions/runner-images/issues/6724).

Allow `__` in profiler names, now that we split on `:`.

`exprelr` was described all wrong, fixed.

Mechanism ids are assigned by iterating through an unordered map within `fvm_mechanism_data`. This changes the container in question to an ordered `std::map`.

See the same jobs fail, which turn out to execute using different Python
versions:
-
https://github.com/arbor-sim/arbor/actions/runs/3618204857/jobs/6097794385
-
https://github.com/arbor-sim/arbor/actions/runs/3630819640/jobs/6141257441

Implementation of stochastic calcium plasticity curves, originally
authored by Sebastian Schmitt @schmitts, and adapted here for Arbor
inclusion.
- python example plasticity_stdp.py
- stochastic mechanism calcium_based_synapse.mod
- tutorial

Closes #1987

* Actually set the compiler to the user provided `--cxx` or the scripts default.
* Have a fallback to `c++` in case `arbor.config()['CXX']` does not exist (typically true for pip installed wheels).
* Verbose mode for `a-b-c`

Switch XML parser to pugixml; one less runtime dependency and one less blocker to WASM via emscripten.

required for making an emscripten build of the arbor python library

* make ARB_MODCC functional again

* move add_subdir

* no modcc/CMakeLists.txt is needed also when external modcc is used.

* review

* Remove deprecated spike_detector.

added conditional variant for cuda builds to enable GPU-based random number generation

- unit tests and some examples build their own mechanism catalogue
- repeated logic for detecting GPUs, passing compile time flags
- leverage the cmake variables and targets which are already there since these catalogues are internal to arbor
- build a static library with mechanisms, which links to arbor-private-deps

Add the VOLTAGE_PROCESS mechanism kind to modcc, allowing for direct writing to the membrane voltage
Obviously these are extremely sharp tools and can break the cable model, so we add some constraints to their
use (see docs). Still, user discretion is required.

As a sneak peek for what this might be useful for

* implementing voltage clamps
* limiting membrane voltage (irritatingly also a kind of clamping)
* faking ABSTRACT_CELL like behaviour (although that might require a bit more work)

Closes #1343

Throw an informative error instead of `bad_optional_access` when we do not find a global or cell default for an ion's diffusivity.

Added common math functions which I came across in the wild quite frequently. Implementations may be more optimized than hand-rolled versions.

- sqrt: usual square root
- step_right: right-continuous heaviside step function
- step_left: left-continuous heaviside step function
- step: heaviside step function with half step at 0
- signum: sign bit function
- extended iexpr with the step_left, step_right variants

fix ornstein_uhlenbeck example on gpu:
- use make_catalogue instead of make_catalogue_standalone
- link to arbor-private-deps
- mark all files as c++ sources when using cuda-clang or hip-clang

- Correct `.gitmodules`
- Update all git submodules to latest released versions, except google-benchmark
- Have CHANGELOG
- Breaking changes since v0.7:
  - A change in API: `arbor.cable_cell` has the labels and decor arguments swapped. I.e.: `(tree, labels, decor)`
-> `(tree, decor, label)`. Labels are now optional.
  - Mechanism ABI version is bumped to 0.3.1. https://github.com/arbor-sim/arbor/pull/1884
  - Remove the `generate-catalogue` script.  `modcc` accepts now a list of NMODL files and is able to spit out a catalogue.cpp file
  - Rename spike detector -> threshold detector
  - Remove access to time `t` in NMODL.
- Major dependency version bumps:
  - GCC: 9 and up
  - CUDA: 11 and up
  - Clang: 10 and up

Addresses #1987

* reproducibility test for random number generator
* C++ example demonstrating Ornstein-Uhlenbeck process
* Python example featuring stochastic mechanism
* stochastic mechanism catalogue

The Python example is contributed by Jannik Luboeinski's (@jlubo): https://github.com/jlubo/arbor_ou_lif_example
and adapted to fit into Arbor.

* fix compiler warnings

* fix some more warnings

Co-authored-by: Brent Huisman <brenthuisman@users.noreply.github.com>

- Add (again) a reset voltage E_R that for V_m = E_R in the refractory period.
- Make V_m decay towards resting potential E_L instead of 0.
- Adjust docs and tests accordingly.
- Double check this against literature.
- Explain weight, add C++ API.

* Added fixture dev docs. Made fixtures more robust.

Co-authored-by: Brent Huisman <brenthuisman@users.noreply.github.com>