Remove use_gpu when data is available

[guyvdbroeck]

We currently need to synchronize between the device used for the bitcircuit, the data, and often a use_gpu flag. Standardize on always using isgpu(data) for this decision whenever the data is available; move the bitcircuit to the same device as the data, and remove the use_gpu flag. Don't repeat infrastructure to move things to the GPU, assume the outside user makes that decision by saying togpu(data) or not.

[khosravipasha]

Also related, we are using reuse for resuing memory for different batches, but not across epochs, might be good to change the high level API to make it easy to do.