For tree-based methods, use min_impurity_split instead of min_samples_*?

In sklearn 0.18, they introduced `min_impurity_split` for tree-based methods, e.g., [random forests](http://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html). Description:

>Threshold for early stopping in tree growth. A node will split if its impurity is above the threshold, otherwise it is a leaf.

This seems like a much better way to choose the early stopping criteria in decision trees, as selecting a number of samples to stop at seems quite arbitrary.

We should evaluate the effect of this parameter, find out some good ranges, and change over TPOT's tree-based methods to this parameter in place of the `min_samples_*` parameters.