CUDA architecture parser accepts unsupported pre-Hopper a suffixes

The exact CUDA architecture parser validates the numeric base code but accepts an a suffix even for architectures that have no architecture-specific target.

Reproduced on main at f9474304bbfe94bf7ffe0f8dbbcc468c1ff0ceb1:

python -c "import build_arch; print(build_arch.parse_cuda_arch_token('8.0a').cubin_flag)"

Actual output:

-gencode=arch=compute_80a,code=sm_80a

The same problem affects 7.0a, 7.5a, 8.6a, 8.7a, and 8.9a, including compact/prefixed spellings and +PTX variants. These targets are absent from NVIDIA's supported architecture list.

Expected behavior: reject unsupported base/suffix combinations with CudaArchError before returning invalid compiler flags, and identify the original input in the error. Valid existing targets, including 9.0a, 10.0a, 10.3a, 12.0a, and 12.1a, should retain their exact compiler flags and explicit +PTX behavior.

The parser is independent of PyTorch, so this can be covered using the existing CPU-only parser tests. No CUDA compilation was performed for this report.

Start with build_arch.parse_cuda_arch_token and the existing CPU-only parser tests. Reproduce the 8.0a example, then cover unsupported pre-Hopper a combinations, compact or prefixed spellings, and +PTX variants. Done means these inputs raise CudaArchError identifying the original input while valid targets such as 9.0a and 10.0a retain their exact flags and +PTX behavior.