Add CUDA memory pool allocation function aliases for MUSA compatibility

[popsiclexu]

• Add CUDA memory pool allocation function aliases (_cuda_beginAllocateCurrentThreadToPool, _cuda_endAllocateToPool, _cuda_releasePool) in the C++ extension, backed by MUSA allocator APIs, to support from torch.cuda.memory import ... usage on MUSA devices
• Implement thread-based filtering in _musa_beginAllocateCurrentThreadToPool to match CUDA semantics (instead of the stream-based fallback from _patch.py)
• Inject these functions into torch.musa.memory at C++ extension load time, so downstream code using CUDA memory pool APIs works transparently on MUSA

[augmentcode]

🤖 Augment PR Summary

Summary: This PR improves CUDA-API compatibility on MUSA by exposing CUDA-named memory pool allocation helpers through torchada’s C++ extension.

Changes:

• Adds C++ bindings for _cuda_beginAllocateCurrentThreadToPool, _cuda_endAllocateToPool, and _cuda_releasePool backed by MUSA caching allocator APIs.
• Implements thread-based filtering for “current thread” semantics when allocating to a pool.
• Injects the new functions into torch.musa.memory when the extension loads so torch.cuda.memory imports work under torchada’s CUDA->MUSA redirection.

Technical Notes: Injection happens after the extension is successfully built/loaded and marked as loaded, and is intended to make downstream CUDA memory pool API calls operate transparently on MUSA.

_{🤖 Was this summary useful? React with 👍 or 👎}

[augmentcode]

Review completed. 2 suggestions posted.

Comment augment review to trigger a new review at any time.