Implement support for categorical variables

[ogrisel]

pygbm should accept pandas.DataFrame instances as in X in gbm.fit(X, y) (as well as predict, in addition to numpy arrays).

This would make it more natural to accept heterogeneously typed features in the training data.

I am not sure what is the best strategy is to deal with categorical variables. Reviewing the strategy implemented in LightGBM and XGBoost would be helpful.