Standard interface for `predict`

[ots22]

At the moment there are very few constraints on a model's predict method (particularly its return type). This is hard for a user (who must figure this out for each model - it isn't recorded in a standard place), and makes switching models tedious.

One idea is to settle on a standard interface for each task, although a drawback is that there may be too many task variants to do this exhaustively.

Another is to record the function signature somewhere (perhaps in .scivision/model.yml - beyond what is there now) and make this available from the library.

We could start by writing down the interfaces used in each example currently in the gallery and try to settle on something for common tasks.

Some more discussion here: https://hackmd.io/14jVOkbtTQOgxtc636kdJw

[ots22]

Potential source of 'tasks' to consider: http://taskonomy.stanford.edu/