Consider `.generate()` allowing tokenized inputs if `preprocessor=None`

[mattdangerw]

We want .generate() to feel similar to .fit(), .predict(), and .evaluate() on a model.

We should consider allowing .generate() to token id input directly (for example, with custom packing of special tokens), if preprocessor=None. In this case the output should also be token id sequences, and not strings.

This would match with .fit(), .predict(), and .evaluate() behavior on a task model when preprocessor=None.