[clang] [clang-linker-wrapper] Pass on --cuda-path to child clang processes (PR #149107)

[Joseph Huber via cfe-commits]

jhuber6 wrote:

It's passed to the linker wrapper because it needs `fatbinary` for CUDA, but it should *also* be forwarded to the embedded clang job. If you compile with `-v` do you not see it on the `clang --target=nvptx64-nvidia-cuda` job?

https://github.com/llvm/llvm-project/pull/149107