`.generate()` should support input datasets

It would be a good idea to add support for passing a dataset to `generate()` for a few reasons.

 - In the seq2seq case, summarizing/translating an entire dataset is a real use case. We should in theory see the same performance gains using `tf.data.Dataset.prefetch` setup as with `fit()` and `predict()`.
 - We want `generate()` to have a similar UX to `fit()` and `predict()`.