[GPU][CUDA] Add support for linear tree with device=gpu, device=cuda

[dragonbra]

Summary

Add support for linear tree with device=gpu, device=cuda

Motivation

Linear tree is a useful feature to improve model's performance. However, it can be only used in CPU version now. I think it would be great if we can add support in GPU and CUDA version

Description

By reading the code, I found that the gpu_tree_learner now calculates the gradients and hessians in GPU device and pass them back to CPU to call the train() function in serial_tree_learner. The linear_tree_learner's training process is based on serial_tree_learner's and used Eigen library, so it should be easy to implement this feature if we do not take GPU acceleration of training linear model leaves into consideration.
I have already successfully tried it in device=gpu version and I will submit a PR when I have time to organize the code.
I haven't done experiments in device=cuda version. It would be great if anyone can contribute to it.

References

Trees with linear models at leaves #3299

LightGBM/src/treelearner/linear_tree_learner.cpp

Line 62 in f8ec57b

Tree* LinearTreeLearner::Train(const score_t* gradients, const score_t *hessians, bool is_first_tree) {