GitLab

* CUDATreeLearner: free GPU memory in destruuctor if any allocated

* Minor changes: checking for num_gpu_feature_groups is not needed

* Trigger CI again