[libc] [llvm] [LLVM] Port 'llvm-gpu-loader' to use LLVMOffload (PR #162739)

[Artem Belevich via llvm-commits]

Artem-B wrote:

> This may require you to change the build of the NVIDIA bot for libc. Ideally you just need to add `offload` to the runtimes list and make sure that `libLLVMOffload.so` and `libcuda.so` exists somewhere in the path.

Are the libraries expected to exist somewhere in the search path on the build host or on the host where I run the tests?

https://github.com/llvm/llvm-project/pull/162739