Reduce redundant GPU allocations

[thorstenhater]

Introduction

• Do not allocate and fill storage for resetting ion concentrations if these are never changed.
  Reasoning: If concentrations are never changed, we do not reset them and thus do not need
  to store the values.
• Remove a duplicate of the CV area in solver and shared state and pass a reference during solving
• Remove redundant solution from GPU solver
• Do not store / allocate diffusion values if not needed.

This saves per CV

• 1 x 8B for cv_area unconditionally
• 1 x 8B for Xd for each ion with no diffusion is in use (majority of cases)
• 2 x 8B for Xi for each ion (reset and init) if not written (reasonably often)
• 2 x 8B for Xo for each ion (reset and init) if not written (majority of cases)
• 1 x 8B for eX reset for each ion if not read (majority)
• 1 x 8B for eX for each ion if not read (rarely)

In my standard benchmark, busyring with complex cells, this saves about 18% of the total GPU
allocation for the cell data (shared_state).

This has become a mixed bag, fixing a few additional things that came up during testing this:

• a bug in event handling on GPU
• pybind11 stub generation

[jlubo]

This is nice!

The previously high memory consumption prevented the execution of my large network application with multi-compartment neurons (see here for the code of the single-compartment case, the code for the multi-compartment case is not public yet) on common GPUs. The present PR reduces the GPU memory consumption for that application from >> 8 GB to ~100 MB.

Also, the whole variety of tests for my network application do pass (for both the single-compartment and the multi-compartment case).

[thorstenhater]

You can also try #2394, which takes the concept one step further.

[thorstenhater]

Spack is failing due to pb11-stubgen not being updated. I made the corresponding PR here
spack/spack#47162

[boeschf]

the conditional initialization and allocation of the different ion state arrays is not very easy to follow. There must be a better way to do this, but that's maybe for another PR.

[boeschf]

can we add a spack variant that would disable the pybind11 stubgen? This would be consistent with the cmake and would allow the spack CI to still pass.

[boeschf]

thanks, looks good to me