hotfix: inplace pin memory caused cudaErrorHostMemoryAlreadyRegistered

[specture724]

Resolve #67
Manually unregister the memory pinned with cudaHostRegister by cudaHostUnregister
Renamed test_pin_memory.py to test_reuse_pin_memory.py

[Copilot]

Pull request overview

This PR fixes a CUDA error (cudaErrorHostMemoryAlreadyRegistered) that occurred when reusing inplace pin memory by implementing manual memory unpinning with cudaHostUnregister. The changes also make inplace pin memory the default behavior and add NPU device compatibility checks.

Key Changes:

• Implemented manual memory unpinning using cudaHostUnregister for CUDA-registered memory to prevent reuse errors
• Changed default value of use_inplace_pin_memory from False to True
• Added NPU device detection to disable inplace pin memory on unsupported hardware

Reviewed changes

Copilot reviewed 3 out of 4 changed files in this pull request and generated 7 comments.

File	Description
checkpoint_engine/ps.py	Added manually_pinned flag to MemoryBuffer, implemented _unpin() function for manual memory unpinning, changed default for use_inplace_pin_memory to True, added NPU device check, and updated documentation
tests/test_reuse_pin_memory.py	New test file (renamed from test_pin_memory.py) that validates shared memory pool registration and unregistration behavior
tests/test_inplace_unpin.py	New test that validates repeated pin/unpin cycles work correctly without CUDA errors
tests/test_update.py	Removed unused variable assignment and changed directory cleanup from shutil.rmtree() to os.removedirs()

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.