Early stopping

[jaberg]

Cleaned up early stopping code, but left it as a modification to the BaseASGD and NaiveASGD.

Rationale: I think early stopping of some description is always desired, and the old behavior can be obtained by setting min_observations = max_observations = n_iterations * N_POINTS.

If now or later someone says that this early stopping heuristic is can be improved, then I suggest just modifying the base class to use the other heuristic. I hesitate to use some elaborate plugin system to manage the early stopping heuristics because if someone wants that kind of control then they should use partial_fit directly. That said, I think a fit() with no stopping heuristic is basically not very useful.