Generate CUDAConfig.h only for CUDA builds (#78218)
This should prevent failures like https://github.com/pytorch/pytorch/pull/77002 from sneaking in as CUDAConfig.h would no longer be available for cpu builds.
Note from 2018 about MIOpen builds do no seems relevant, though CUDAConfig.h is still needed by ROCm (tested in https://github.com/pytorch/pytorch/runs/6613660811?check_suite_focus=true build)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78218
Approved by: https://github.com/seemethere, https://github.com/atalman
