Properly cache the results of CUDA native detection.

I found some issues when using VTK-m from inside VTK. The issues where that on reconfigures the device architectures flags where dropped by the caching mechanism.