Fixed problem with how DecisionTree handles ordered categorical features.

Background: Ordering refers to how nodes can split on categorical features.  See the docs for more info.

Previously, it chose a fixed ordering a priori.
Now, it chooses a different ordering for each node, using the aggregated stats.

For binary classification and regression, this is a bug fix.
For multiclass classification, this improves a heuristic.

Related changes:
* splits and bins are no longer pre-computed for ordered categorical features since they are not needed.

Author: jkbradley