[CUDA] Enable use of subset Dataset in new CUDA version

[shiyu1994]

Summary

In #4630, we did compress the Dataset memory in CUDA even when the sampling rate is very small. While this is done in CPU version,

LightGBM/src/boosting/gbdt.cpp

Lines 253 to 262 in d4cdbcf

	if (!is_use_subset_) {
	tree_learner_->SetBaggingData(nullptr, bag_data_indices_.data(), bag_data_cnt_);
	} else {
	// get subset
	tmp_subset_->ReSize(bag_data_cnt_);
	tmp_subset_->CopySubrow(train_data_, bag_data_indices_.data(),
	bag_data_cnt_, false);
	tree_learner_->SetBaggingData(tmp_subset_.get(), bag_data_indices_.data(),
	bag_data_cnt_);
	}

Motivation

Enable subset Dataset (physically compress Dataset in memory when low sampling rate) can speedup the training with GOSS or Bagging.

References

A related discussion, #4630 (comment)