Allow usage of generator-like objects instead of numpy arrays for training

By using a single `numpy` array at train time, the user is currently limited to building models on dataset sizes that fit in RAM. Instead, we should allow the user to train on a generator (akin to `fit_generator` in keras), which can read data as needed from disk. Perhaps the user can pass a `pytorch.Dataset` instead of a numpy array.