Don't run nvcc native detection multiple times, as the results are cached.

This fixes a issue where when CUDA is enabled and VTKm_CUDA_Architecture is set to native, we would run nvcc each time somebody explicitly enabled the CUDA component ( each example ).