second attempt add support for tf Dataset

[david-thrower]

The problem

• Unfortunately, we can't use a separate dataset for the x and y args for tf.keras.Model.fit().
• This makes us unable to run this on out of memory generators (like the generator that streams the auto-regressive expansion of tokenized text).
• This precludes training the LLM on a reasonable sized dataset ... unless we want to pay for a TB of RAM...

The feature:

Allow x=foo and y=bar to be passed or dataset=Dataset() # that pachages foo and bar as x and y ... and selectively run .fit based on the presence or absence of these mutex param configurations.