Improve source build robustness

[liuyao0322]

Summary

• Use the invoking Python interpreter for CMake/PyTorch discovery.
• Match PyTorch's CXX11 ABI and add clearer build diagnostics.
• Make CUDA architectures and build parallelism configurable.
• Document toolchain troubleshooting for PyTorch header failures.

Validation

• Passed python3 -m py_compile setup.py
• Passed git diff --check
• Passed source extension build with CUDA 12.8 and PyTorch 2.10.0+cu128:
  python setup.py build_ext --inplace
• Passed wheel build with CUDA 12.8 and PyTorch 2.10.0+cu128:
  MAX_JOBS=8 python setup.py bdist_wheel