Use __pyx_capi__ for CUDA driver function pointers

[Andy-Jost]

Summary

Replace the _CUDA_DRIVER_API_V1 capsule with direct extraction of function pointers from cuda.bindings.cydriver.__pyx_capi__ at module import time.

Changes

• Remove CudaDriverApiV1 struct and _get_cuda_driver_api_v1_capsule() from _resource_handles.pyx
• Remove load_driver_api(), ensure_driver_loaded(), and related machinery from resource_handles.cpp
• Add extern declarations for driver function pointers in resource_handles.hpp
• Populate function pointers at module import using PyCapsule_GetPointer() with PyCapsule_GetName() for signature lookup
• Update DESIGN.md to reflect the new __pyx_capi__ approach

This simplifies the architecture by eliminating the custom capsule struct and its loading machinery. The driver function pointers are now populated directly from Cython's built-in cross-module API mechanism.

Stats: 4 files changed, 228 insertions(+), 392 deletions(-) (net -164 lines)

Test Plan

• Build succeeds
• Import works on CPU-only machines (no GPU/driver required)
• test_stream.py, test_event.py, test_memory.py, test_device.py pass
• CI tests pass

Closes #1450

[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.

[Andy-Jost]

/ok to test 3921ca6

[rwgk]

Awesome! I didn't find anything visually, or with Cursor.

If we had more CI resources we'd be flying! :-)

[kkraus14]

Instead of reaching for the __pyx_capi__ which is an implementation detail of Cython, did we try using cdef api and cdef public? https://cython.readthedocs.io/en/latest/src/userguide/external_C_code.html#c-api-declarations

This generates header file(s) that we can use nicely as expected instead of reaching into internal structs generated by Cython.

[Andy-Jost]

/ok to test 0b08aa4

[github-actions]

Doc Preview CI
🚀 View preview at https://nvidia.github.io/cuda-python/pr-preview/pr-1466/
https://nvidia.github.io/cuda-python/pr-preview/pr-1466/cuda-core/
https://nvidia.github.io/cuda-python/pr-preview/pr-1466/cuda-bindings/
https://nvidia.github.io/cuda-python/pr-preview/pr-1466/cuda-pathfinder/
Preview will be ready when the GitHub Pages deployment is complete.

[Andy-Jost]

@kkraus14 No, I didn't consider cdef api.

We discussed __pyx_capi__ in a review meeting and concluded it was stable based on this discussion.

If we were to use cdef api instead, it would create a build-time dependency on cuda-bindings:

• We'd need to modify the cuda-bindings code generator templates to add the api keyword to the driver function definitions
• Cython would then generate *_api.h headers that we'd need to export and ship with cuda-bindings
• cuda-core would #include these headers and call an import_*() function at initialization

The tradeoff is that cdef api is an official Cython feature, whereas __pyx_capi__ is technically internal (though stable in practice).

If you'd like me to look into using cdef api, please let me know and I will create a follow-up issue.

[leofang]

Wonderful! Thanks for looking into it @Andy-Jost.

[leofang]

Sorry, for some reason I missed the cdef api discussion. That's something we'd like to explore (#464) but I don't think we need to tie it to this safety feature.