Add a Python evaluator?

[frostedoyster]

eval is great to have, but users often want to do their own evaluation of the model, with complicated error metrics, splitting of the test set into different subset, etc
Given the common interface of models after exporting, it would be relatively simple to provide a Python evaluator that the user can call to do custom evaluation and that we could also use internally for eval. I think a small tutorial would be the best way to expose it to the user