Add `impurity` field in Treelite model representation

Right now, Treelite model object only keeps `gain` but discards `impurity` information from scikit-learn random forests.

Start by tracing Treelite's model representation and the scikit-learn random-forest import path, focusing on where gain is currently retained and impurity is discarded. Done means impurity information from scikit-learn random forests is preserved in the Treelite model representation and existing relevant checks still pass.

Written by the indexing model from the issue text.