ENH: Validate output units and outputs in multi-output models

[adriangb]

Cross ref #143 (comment).

We currently check the number of Keras outputs, but do not check the shapes/units in those outputs, and do not validate any multi-output models.

• Do we want more validation here? Are we okay with the added introspection?
• If so, what is the source of truth and what is being validated?