Improve awkward handling of GMX_CUDA_TARGET* help