[llvm] [OpenMP] Port the OpenMP device runtime to direct C++ compilation (PR #123673)

[Joseph Huber via llvm-commits]

jhuber6 wrote:

> I tested this patch with the CMake cache file in `offload/cmake/caches/AMDGPUBot.cmake` and see errors about some nvptx that is built. The cache only enables `host;AMDGPU`. Does this patch miss / remove guard?

I didn't see issues locally, if you run `ninja check-offload` there's always like 12 NVPTX tests that fail for me with and without this patch. Could you copy them here?

https://github.com/llvm/llvm-project/pull/123673