Move some functions to GPU

[caseyjlaw]

Check out https://cupy.chainer.org. Good mapping of numpy to cuda functions. Flexible kernel design syntax, too.
Good plan could be to move some of prep functions in such that data are put in GPU during data_prep call.
First, need to investigate how many moves in/out of memory would be required. If we use cupy and rfgpu, would data need to come out and then in again to GPU?