Data expectation for primitives that support continue_fit.

As we discussed, it's not clear who has the responsibility of shaping the data that is going to be pass to a primitive via set_training_data(). For the regular case when fit is called, is trivial and straightforward, but for the cases of the continue_fit is not clear, since we can have cases where not all classes (for classification) are passed on a given batch and could break the primitive when continue_fit is called.

So, the question is, who is responsible of handling these cases, the primitive that implements continue_fit? The way that we are implementing batching? Another primitive?