Add label leakage check after training models in AutoML

Per @kmax12's comment in #917, we could do a label leakage check after training a model, checking if it scored very highly but only has a single feature with all the importance. This could indicate that the feature that scored highly is correlated to the target.

This would need a little more design: currently, we only perform any data checks before searching. How should we support data checks (or only some data checks) running after a batch / iteration?