Investigate eliminating _CXX_API capsule for resource handle functions #1452

Summary

PR #1368 uses a PyCapsule mechanism (_CXX_API) to share resource handle functions across Cython modules. This issue is to investigate whether this capsule can be eliminated entirely.

Current Architecture

The _CXX_API capsule exports function pointers from _resource_handles.so to other Cython modules (_buffer.pyx, _stream.pyx, _event.pyx, etc.). The stated rationale is to avoid duplicate static/thread-local state if multiple modules linked the C++ code directly.

Proposed Investigation

Investigate whether Cython modules can call directly into the resource handles code without the capsule indirection:

• If resource handle functions are regular cdef functions (non-inline) in _resource_handles.pyx, can other modules cimport and call them directly?
• Would this approach correctly share static state through _resource_handles.so?
• Are there performance implications for nogil code paths?

Background

When module A cimports a cdef function from module B, the call goes through B.so at runtime via Python's import mechanism. This should mean static state in B stays in one place, potentially making the capsule unnecessary for pure Cython-to-Cython usage.

References

• PR RAII resource handles for CUDA objects #1368 - RAII resource handles implementation
• Issue Investigate using __pyx_capi__ to simplify resource handle architecture #1450 - Related investigation for cuda.bindings driver API capsule

Goal

Simplify the architecture by removing the capsule indirection if it's not required, reducing complexity while preserving correct behavior.