[Inductor CI] Use string format for cuda-arch-list input to prevent 8.0/9.0/10.0 etc from being interpreted as 8/9/10 (#89279)
Currently or in future whenever we change the cuda-arch-list to num.0, github action or some agent would pass just num to TORCH_CUDA_ARCH_LIST
This num is not regex matched during cuda arch analysis phase. (here: https://github.com/pytorch/pytorch/blob/c5fafb4e1694f141d8a1a31142cce4049d9057ed/cmake/Modules_CUDA_fix/upstream/FindCUDA/select_compute_arch.cmake#L229)
Example failure: https://github.com/weiwangmeta/pytorch/actions/runs/3495656108/jobs/5852735299
Unknown CUDA Architecture Name 8 in CUDA_SELECT_NVCC_ARCH_FLAGS
This change reminds us to use e.g. '8.0', '9.0', '10.0' etc instead of 8.0, 9.0, 10.0 as GHA or some other agent may erroneously truncate it to pure numbers.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89279
Approved by: https://github.com/desertfire, https://github.com/atalman
