Split models into a trainer and model pair

[sgoldenCS]

To reduce the reliance on one large model class, we will try to split the training from the model architecture. This works well in PyTorch, but doesn't match standard Keras models. This branch is dedicated to creating separation between the training (and its configuration including optimizers) and the model itself.

It is possible that this idea will not be successful, or it may be impractical for Keras models. But we can't know without testing it.