GPU detection fails when CUDA compilation requires CUDA_HOST_COMPILER to be set

[Sven Verdoolaege]

I don't seem to have permission to write to the repo, so I'm attaching the patch here.
Without it, I get "Automatic GPU detection failed. Building for common architectures.".

0001-use-CUDA_HOST_COMPILER-to-detect-installed-GPU.patch.txt

[Priya Goyal]

this has been merged in pytorch/aten master and can be closed now :)