[transfer-engine] Ship intra-node NVLink transport in x86_64 CUDA wheels (#3547)

* [transfer-engine] Ship intra-node NVLink transport in x86_64 CUDA wheels

Build the x86_64 cuda and cuda13 wheels with -DUSE_INTRA_NVLINK=ON so the
nvlink_intra transport is available to prebuilt wheel users, and keep the
RDMA/TCP fallback correct for builds that enable USE_INTRA_NVLINK without
USE_MNNVL.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

* Modify PyTorch test configuration in CI workflow

Updated the test name for PyTorch to reflect versioning and removed specific version installation.

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Xun Sun <UNIDY2002@outlook.com>