CUDA: Make sure CUDA architectures are in order

Make sure we always group major versions together instead
of interleaving 100 and 120 families.