fix: protect dependency headers and align stable ABI versions

[WindDaoist]

Summary

• Keep dependency include paths available to the compiler while excluding configured dependency roots from in-place CUDA-to-MUSA porting.
• Add TORCHADA_EXCLUDE_DIRS, accepting paths or complete directory names such as torch_musa; explicitly supplied source directories remain portable.
• Select stable-ABI compatibility from the installed torch version: torch 2.9 uses torchada's header backport for MUSA extension builds, while torch 2.11 and newer use the native stable ABI without patching installed headers.
• Align the stable MUSA stream and BLAS-handle shims and mappings across the torch 2.9 compatibility path and the torch 2.11 native path.
• Extend the existing end-to-end stable-ABI pytest with vLLM's weak_ref_tensor surface and version-specific backport/native assertions.

Motivation

Editable dependencies such as /home/torch_musa can appear in downstream extension include paths. TORCHADA_EXCLUDE_DIRS=torch_musa keeps those dependency headers available for compilation without treating the checkout as a SimplePorting source root.

The stable header backport is required by the supported torch 2.9 line. Torch 2.11 provides the stable ABI natively, so torchada detects torch.__version__ and skips the on-disk header backport on torch 2.11 and newer. The build remains automatic for downstream extensions; no per-extension opt-in is introduced.

Tests

Targeted unit coverage:

51 passed, 1 skipped, 204 deselected

PYTHONPATH=src python -m pytest -q \
  tests/test_cpp_extension.py \
  tests/test_inplace_porting.py \
  tests/test_cuda_patching.py \
  -k 'stable or dependency or inplace' \
  tests/test_stable_abi_mappings.py

End-to-end stable-ABI matrix on one MTT S5000 with MUSA driver/runtime 5.2.0:

torch 2.9.1.post1+musa5.2.0
torch_musa 2.9.1.post1+musa5.2.0
test_stable_abi_ops_with_torch29_backport: 1 passed in 16.50s

torch 2.11.0.post1+musa5.2.0
torch_musa 2.11.0.post1+musa5.2.0
test_stable_abi_ops_with_native_torch211_abi: 1 passed in 16.28s

The same vLLM-derived weak_ref_tensor operation is compiled and executed in both environments. The torch 2.9 case asserts that the compatibility backport runs; the torch 2.11 case fails if torchada attempts the backport and compiles without torchada's stable compatibility include or force-included boxer header.

Full torchada pytest suite on commit 2b846dd, using the same MTT S5000 and upgrading torch/torch_musa in place between runs:

python -m pytest -q

torch 2.9.1.post1+musa5.2.0 / torch_musa 2.9.1.post1+musa5.2.0
430 passed, 19 skipped, 3 warnings in 54.69s

torch 2.11.0.post1+musa5.2.0 / torch_musa 2.11.0.post1+musa5.2.0
430 passed, 19 skipped, 5 warnings in 55.35s

The additional warnings on torch 2.11 are profiler and TorchScript deprecation warnings; both runs completed without test failures. Downstream vLLM-MUSA build and serving validation is intentionally left to a separate integration effort.

[yeahdongcn]

@Joey-gvwal Please take a look. Thanks!