Fix primary context cleanup during Python shutdown

[bdice]

Summary

• avoid calling the CUDA Bindings Cython wrapper from the primary-context TLS destructor after CPython has finalized
• preserve cuDevicePrimaryCtxRelease during normal thread cleanup while leaving process-exit cleanup to CUDA/the OS
• add a subprocess regression test covering Device().set_current() followed by interpreter shutdown

Fixes #2743.

Root cause

get_primary_context() caches a ContextHandle in thread-local storage. When the main thread's TLS destructors run after Py_Finalize(), the handle deleter calls p_cuDevicePrimaryCtxRelease. That pointer targets cuda.bindings.cydriver.__pyx_capi__["cuDevicePrimaryCtxRelease"], whose generated Cython entry path touches Python thread state. The result is a post-atexit segfault in PyThreadState_New.

The existing GILReleaseGuard avoids manipulating the GIL after finalization, but the Cython driver wrapper invoked immediately afterward has the same Python-runtime requirement. The new guard skips that wrapper only when Python is no longer usable.

Testing

• pixi run --manifest-path cuda_core -e cu12 python -m pytest cuda_core/tests/test_device.py -q
  • 166 passed, 2 skipped
• minimal shutdown probe exits 0:
  • python -c 'from cuda.core import Device; Device().set_current(); print("initialized")'
• repository pre-commit hooks pass

[bdice]

I do not have permission to set metadata on the upstream repository. Suggested PR metadata: labels bug and cuda.core; milestone cuda.core 1.2.0.

[bdice]

This test will only reproduce the failure I saw if it is run with cuda-bindings <=12.9.2 or cuda-bindings >=13.0.0,<=13.0.1. However, it is good to include to prevent regression.

[seberg]

Thanks Bradley.

Surprising to me that these are cleaned up. The pattern is common, either way.

(Whether this should PrintUnraisable is a seperate issue.)

[github-actions]

Backport failed for 12.9.x, because it was unable to cherry-pick the commit(s).

Please cherry-pick the changes locally and resolve any conflicts.

git fetch origin 12.9.x
git worktree add -d .worktree/backport-2744-to-12.9.x origin/12.9.x
cd .worktree/backport-2744-to-12.9.x
git switch --create backport-2744-to-12.9.x
git cherry-pick -x e74d4d557d6a436d5064c71480a23c3837699f41

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.

[seberg]

Ah, nvm. This code didn't exist at all in 12.9, so no backport.

[leofang]

@Andy-Jost Would be nice to check in which PR we added this. I am 100% certain that this goes against cuda.core assumptions. We can't afford primary context from being destroyed or reset. Same as in cccl-rt, we cache a ton of things from the primary context, which would become dangling ptrs if the context is gone.