Add critical section guards to accessors

[seberg]

Protect accessors with cython critical sections so free-threaded execution avoids races when reading cached state.

@cython.critical_section effectively takes a lock based on the first argument passed in. (Although that lock can be released to avoid deadlocks, it is as safe or safer as the GIL, but that doesn't make it 100% in all situations.)

At least one of these was noticed by free-threaded test runs. I am not certain that they will guarantee identity returns, but pretty sure this will make them thread-safe.
(The reason being that at least on non free-threaded builds they may release the GIL or, on free-threaded, do API calls that release the critical section temporarily.)

About the .pyi fixes: I think these are incorrect but cython is already used in other places and shouldn't be a runtime/typing requirement at all.

Towards gh-2194 which will add some testing although not strictly guaranteed for all paths covered here.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[seberg]

/ok to test 652a3e7

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-2215/
https://nvidia.github.io/cuda-python/pr-preview/pr-2215/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-2215/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-2215/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[mdboom]

About the .pyi fixes: I think these are incorrect but cython is already used in other places and shouldn't be a runtime/typing requirement at all.

We probably need to submit a fix to stubgen-pyx to not do this. I think it's probably just being "literal" about what's there. I don't think it "breaks" our .pyi files, just unnecessary.

There are obviously a lot more properties/accessors in the codebase. Besides seeing multi-threaded test failures, are there other criteria we should follow about when to apply these decorators?

[seberg]

We probably need to submit a fix to stubgen-pyx to not do this. I think it's probably just being "literal" about what's

Yeah, right now it seems you need cython and will just get Any (because Cython isn't fully typed I think). A simple "ignore these" feature would be good.

There are obviously a lot more properties/accessors in the codebase.

Whenever we mutate a cdef object (global or class instance) then there can be a race that breaks the INCREF/DECREF balancing.
That is even if we do not care about calls being strictly idempotent (once cached a cache value can't be replaced). And that incref/decref balancing issue is strictly a free-threading issue.

I would hope that Cython will provide things to make this easier in the future (maybe atomic object).

You are not wrong, in that I should make another pass and maybe TSAN testing could flush out some as well. In practice hitting these issues is very unlikely but if you do...