test(pinned): use mapped memory for UVA pointer check

[huo-ju]

Summary

• use alloc_pinned_tensor in test_host_device_ptr_is_identity_under_uva
• verify UVA pointer identity with mapped pinned memory, which is the supported host_device_ptr contract
• update the test comment to reflect CUDA 13 registration validation

Why

On Linux/UVA with CUDA 13, cudaHostGetDevicePointer rejects an unregistered pageable host pointer with cudaErrorInvalidValue. The existing test passes a pageable tensor, contradicting the pinned + mapped contract of host_device_ptr and failing before pointer identity can be checked.

This is a test-only change; runtime code is unchanged.

Testing

Environment:

• OS: Ubuntu 26.04 LTS, x86_64
• CPU: Intel Core i9-9920X
• GPU: 4 × NVIDIA TITAN RTX 24 GiB (SM75); test executed on the default CUDA device
• NVIDIA driver: 610.57.04
• Python: 3.14.4
• PyTorch: 2.12.0+cu130
• CUDA runtime: 13.0
• CUDA toolkit: 13.3 (nvcc V13.3.73)
• Checkpoint: N/A; these tests do not load model weights

Commands:

pytest -q tests/kernels/test_pinned_tensor.py::test_host_device_ptr_is_identity_under_uva
pytest -q tests/kernels/test_pinned_tensor.py

Results:

• origin/main at bd372b6: targeted test failed with cudaHostGetDevicePointer failed ... invalid argument
• this branch at cb44bde: targeted test passed (1 passed)
• this branch at cb44bde: complete pinned-memory test file passed (8 passed)