Change Dataset::CopySubrow from group-wise to column-wise

[shiyu1994]

This is to fix the efficiency problem reported by #3637. When bagging is used, Dataset::CopySubrow multi-threads by feature groups, which is not quite reasonable when a large multi-value feature group exists.

This PR changes the group-wise multi-threading into column-wise multi-threading.

[guolinke]

any benchmark numbers?

[shiyu1994]

Set bagging_fraction=0.4 and bagging_freq=1, and test on the large datasets. This PR speedups the training significantly in MS LTR and Allstate. No obvious effect is observed in other datasets. The metric values are identical with the master, except for MS LTR. I think this PR is ready for merge. But I think we need look into why MS LTR produces 4 different results.

Dataset	row-wise-master time	row-wise-this-pr time	row-wise speedup	col-wise-master time	col-wise-this-pr time	col-wise speedup
higgs	124.89±0.50	124.96±2.16	1.00	152.19±0.69	151.04±0.69	1.01
yahoo	54.41±0.36	54.32±0.55	1.00	62.78±0.75	62.21±0.35	1.01
msltr	190.29±0.85	106.66±0.68	1.78	202.39±4.46	117.24±0.70	1.73
dataexpo_onehot	103.52±0.43	103.70±0.67	1.00	123.75±5.09	120.02±1.57	1.03
allstate	276.00±2.57	204.78±0.51	1.35	307.40±3.08	234.90±2.20	1.31
year	37.49±0.59	37.21±0.50	1.01	17.10±0.22	17.05±0.24	1.00

Dataset	row-wise-master metric	row-wise-this-pr metric	col-wise-master metric	col-wise-this-pr metric
higgs	0.847046	0.847046	0.847046	0.847046
yahoo	0.746130	0.746130	0.746130	0.746130
msltr	0.501659	0.501729	0.501306	0.501659
dataexpo_onehot	0.774887	0.774887	0.774887	0.774887
allstate	0.609139	0.609139	0.609139	0.609139
year	9.010180	9.010180	9.010180	9.010180

[github-actions]

This pull request has been automatically locked since there has not been any recent activity since it was closed. To start a new related discussion, open a new issue at https://github.com/microsoft/LightGBM/issues including a reference to this.