Multiple epochs for Incremental

Incremental should have a `max_iter` hyperparameter to control the number of passes (epochs) over the training data.

Implementing a `max_iter` is relatively straightforward. What we really want is some kind of "as many epochs as necessary until the underlying estimator has converged" parameter, which seems relatively difficult.

Split from https://github.com/dask/dask-ml/pull/258

Start by locating the Incremental entry point and its existing training loop and tests. Review the context from the split-off dask/dask-ml#258 work, then determine how max_iter should control passes over the training data; completion should include coverage showing the requested number of epochs is honored.

Written by the indexing model from the issue text.