Correct CUDA fallback to "fermi" when trying to do auto detection.

The logic would properly fall to fermi, but not set the proper NVCC flags to actually compile targeting the fermi compute architecture.