[OpenMP] Clean up AMD handling for `-fopenmp-targets=amdgcn` arch inference

[OpenMP] Clean up AMD handling for `-fopenmp-targets=amdgcn` arch inference

Previously we had some special handling here that errored out if
multiple architectures were detected. This isn't a problem anymore as
the runtime can handle multi-archicture binaries automatically. So it's
safe to simply take the first architecture that we know works. If users
use `--offload-arch=native` instead it will build for all the
architectures at the same time rather than just picking one. This patch
makes it consisten with the NVPTX version.

Reviewed By: jdoerfert

Differential Revision: https://reviews.llvm.org/D142138