Summary

#4630 implements a new version of CUDA tree learner which dispatch the whole training process of a tree to GPU.

Interaction constraints limit which features can be used in the same root-to-leaf path the decision trees in GBDT. It can be used to integrate prior knowledge about the dataset. It was first introduced into LightGBM by #3126

Motivation

For similar reason with monotonic constraints in #4781, we need to support interaction constraints for the new CUDA version.