Fix build vllm kernel for the first time may encounter failure

[zichen-kuuga]

The vLLM kernel uses header file references with relative paths that need to be modified from csrc/xxx to csrc/xxx_musa. If the default directory order is used, the operation that copies csrc to csrc_musa may execute before changing csrc/xxx to csrc/xxx_musa, causing header files to be missing during the initial installation. To avoid this, we sort directories by depth (deepest first).

All tests in tests/test_cpp_extension.py and tests/test_extension_build.py have passed.

[augmentcode]

🤖 Augment PR Summary

Summary: Fixes first-time vLLM kernel extension builds by ensuring CUDA source directories are ported in a deterministic order (deepest paths first) so header references resolve correctly.
Also: Bumps torchada version to 0.1.49 across package metadata and docs.

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

[augmentcode]

Review completed. 1 suggestion posted.

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