Add support for using the CUDA Toolkit's NVCC to compile HIP code. Fixes: #25143
For now, require the value to be `amd`, since that is the only platform we currently support.