Create `fit` API with better user control

[davidslater]

Currently, there are limited controls for training models in Armory.

Ideally, this would be something that Armory would call out to, provide a PyTorch, TF, or Numpy (or other) dataset, potentially with preprocessing applied, and enable the user full control of model training. They could then return a fit model. They could also use this to pass in various kwargs as desired from the config. Armory would still be in control of what dataset / split is passed in, but the remaining work could be left up to the user.

[lcadalzo]

The current approach in Armory scenario code is essentially as follows:

    def fit(self):
         self.model.fit_generator(self.train_dataset, **self.fit_kwargs)

which uses the fit_generator() API that ART adheres to. This doesn't seem too far off the description you've provided: Armory controls the data, user fit kwargs are passed in from the config, and users can define fit_generator() as desired. What sort of additions/refactors/etc. are you envisioning would enhance user control?