fix: accept PyTorch 2.11 extension path query arguments

[yeahdongcn]

Summary

Support modern PyTorch extension path discovery without dropping legacy CUDA-call compatibility:

• Accept include_paths(device_type, torch_include_dirs).
• Accept library_paths(device_type, torch_include_dirs, cross_target_platform) including keyword calls.
• Retain legacy positional CUDA booleans, (cuda_bool, device_name), cuda=..., device-name precedence and the legacy default CPU library result.
• Respect explicit Torch-path inclusion/exclusion and forward supported options on non-MUSA Torch versions.
• Reject unsupported MUSA cross-target discovery explicitly.

This fixes a real Inductor CPU-helper compilation failure during regional MUSA Graph bring-up: include_paths("cpu", False) was bound to (cuda, device_type), causing False.lower().

Tests

• Changed-file pre-commit hooks: pass.
• 17 CPU-only contract cases: pass (mocked MUSA and old/new native API signatures).
• Cold-cache CPU Inductor compilation plus numerical check on torch/torch_musa 2.11.0.post1+musa5.2.0: pass using the full-Torch CPU helper ABI.

The runtime test applied only these path-helper definitions to the existing torchada 0.1.83 image, preserving its other patches and tuning configs. The test used a fresh TORCHINDUCTOR_CACHE_DIR.

Scope / remaining validation

The tested MUSA build still emits the full Torch CPU prefix, so its first-build smoke requires torch._inductor.config.aot_inductor.link_libtorch = None before compilation. The default header-only mode exposed a separate missing-header issue after this signature error was fixed. This PR does not change that provider/compiler behavior or claim the default header-only mode now works.

Full MAGI-2 MUSA Graph validation is ongoing. No attention, MoE kernel, model quantization, or serving parallelism changes are included here.