fix: preserve CUDA architecture variant suffixes

[adenzhou1350]

Summary

• accept PyTorch-supported CUDA architecture variants such as 9.0a and 12.0a
• preserve the variant suffix through sorting, deduplication, TORCH_CUDA_ARCH_LIST, and emitted -gencode flags
• add regression coverage for JIT and non-JIT paths, including mixed PTX variants

PyTorch 2.11 accepts 9.0a, 10.0a, 11.0a, 12.0a, and related targets. DeepSpeed currently parses the minor component with int(), so these valid values fail before compilation.

Testing

• pytest tests/unit/ops/test_op_builder.py -q -k "not forked_child" — 17 passed
• repository pre-commit hooks for the changed files — all applicable hooks pass
• flake8 under Python 3.11 — pass
• git diff --check — pass

The omitted fork/CUDA integration test is platform-specific and cannot run on Windows; it is unrelated to architecture parsing.