Extract histogram builder from CPU Hist.

[trivialfis]

• Fixes incorrect tests where distributed methods are not actually tested.
• Extract histogram builder to be a reusable module.

Perf

	Dataset	Train time	Accuracy
Before	HIGGS	191.195842235989	0.7342790909090909
	covtype	74.0480548529886	0.9398466476769103
After	HIGGS	184.16188054496888	0.7342790909090909
	covtype	68.06914013501955	0.9398466476769103

[trivialfis]

@ShvetsKS

[codecov-commenter]

Codecov Report

Merging #7152 (16c3d11) into master (f1a4a1a) will increase coverage by 0.17%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #7152      +/-   ##
==========================================
+ Coverage   82.19%   82.37%   +0.17%     
==========================================
  Files          13       13              
  Lines        3999     4017      +18     
==========================================
+ Hits         3287     3309      +22     
+ Misses        712      708       -4     

Impacted Files	Coverage Δ
python-package/xgboost/sklearn.py	89.37% <0.00%> (-0.40%)	⬇️
python-package/xgboost/core.py	83.93% <0.00%> (ø)
python-package/xgboost/dask.py	82.33% <0.00%> (ø)
python-package/xgboost/callback.py	82.20% <0.00%> (ø)
python-package/xgboost/training.py	95.34% <0.00%> (+0.04%)	⬆️
python-package/xgboost/data.py	70.01% <0.00%> (+2.13%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f1a4a1a...16c3d11. Read the comment docs.

[ShvetsKS]

@trivialfis sorry for long response.
LGTM
I didn't observe any degradations also
Thanks for great refactoring