Native support for categorical splits

[RAMitchell]

Existing GBM implementations present two approaches:

1. Implement customised categorical split logic in the tree construction. Sort gradients/hessians of all categories within a feature and find the optimal split. LightGBM/XGBoost follow this approach.

2. Preprocess the dataset to encode new features carrying category information. Catboost uses this approach. It is also implemented here: https://contrib.scikit-learn.org/category_encoders/catboost.html

1 is problematic in that it creates bias towards selecting categorical features over numeric features. Our implementation should ideally be unbiased. Perhaps there is a way to randomly select a single split, like we do for numeric features, in an unbiased way?

2 is problematic in that it potentially involves expensive distributed shuffling or data movement and also requires making a copy of the input matrix. 1) would be significantly less memory hungry.