[gboost] extend table weak learner to group discrete feature values

[accosmin]

• for both single-label and multi-label, group the set of possible discrete feature values into K clusters
• redesign the decision tree to always use a 2-split table weak learner
• the discrete step weak learner becomes the special case of K = 1

[accosmin]

• the decision tree will use only decision stumps for splits