[clang] [llvm] [OpenMP] Remove 'libomptarget.devicertl.a' fatbinary and use static library (PR #126143)

[Joseph Huber via llvm-commits]

jhuber6 wrote:

> How did you disable it? Perhaps it's failing because of the specific error:
> 
> ```
> $ nvptx-arch                                                                                                            
> Failed to 'dlopen' libcuda.so.1
> ```
> 
> For comparison, `amdgpu-arch` doesn't print anything and exits with 0 status (`nvptx-arch` yields 1).

I just set `CUDA_VISIBLE_DEVICES=''` to simulate your machine without a GPU. It's confusing here, because this is direct compilation. It should only be invoking `nvptx-arch` if the user passed `-march=native`. So, doing this for some random file doesn't work for you?
```
$ clang foo.c --target=nvptx64-nvidia-cuda -flto -r -Wl,--lto-emit-llvm -o out.bc
```

> For comparison, `amdgpu-arch` doesn't print anything and exits with 0 status (`nvptx-arch` yields 1).

Can't decide if we should indicate why it failed, maybe add a verbose mode or something.

https://github.com/llvm/llvm-project/pull/126143