Investigate using __pyx_capi__ to simplify resource handle architecture

[Andy-Jost]

Summary

PR #1368 introduces a PyCapsule-based architecture for resource handles. This issue is to investigate whether Cython's __pyx_capi__ mechanism could simplify or replace this approach.

Current Architecture

PR #1368 uses a PyCapsule mechanism _CUDA_DRIVER_API_V1 to pass CUDA driver function pointers from cuda.bindings to the C++ code

Proposed Investigation

Investigate whether __pyx_capi__ (Cython's built-in mechanism for exposing C-level APIs) could replace this capsule:

• Can cuda.bindings expose driver functions via __pyx_capi__ instead of requiring a custom capsule?

References

• Questions about __pyx_capi__ use/stability. cython/cython#3077 - Discussion of __pyx_capi__ for cross-module C API access
• PR RAII resource handles for CUDA objects #1368 - RAII resource handles implementation

Goal

Simplify the architecture if possible, reducing complexity and maintenance burden while preserving the core benefits of centralized resource lifetime management.