[CUDA] Monotonic Constraints on New CUDA Version

[shiyu1994]

Summary

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

Monotonic constraint is a very useful feature during training, which restricts how the prediction values change (monotonically increase or decrease) with some of the feature values. The source code for monotonic constraints can be found in src/treelearner/monotone_constraints.hpp, which is called by SerialTreeLearner.

We found that implement monotonic constraint on our new CUDA version requires more efforts than other light-weight features like extremely random tree and interaction constraints. So I think we should have a separate feature request for this.

References

#2305 (comment) is the initial PR current monotonic constraints in LightGBM. The PDF file attached in the description of the PR is quite valuable for understanding how the monotonic constraints work.