Move GHistIndex into DMatrix.

[trivialfis]

See #4354 for details. This PR moves GHistIndex, we also need to move column matrix in future PRs. Also related: #5143 , #6148

Few more bonus:

• We can remove the is_gmat_initialized_ in hist builder.
• We can implement external memory support if desired.
• We can reuse it in approx tree method.

I will update with benchmarks later.

@SmirnovEgorRu @ShvetsKS

[trivialfis]

Here is a quick benchmark, I think there's no performance difference since I'm just moving the code around:

	master	PR
HIGGS	181.0964948719969	180.65688620665847
Covtype	70.7731420363416	70.98095073600416

The result is from running https://github.com/NVIDIA/gbm-bench 3 times for each dataset.

[ShvetsKS]

@trivialfis sorry for long response.
I also remeasured benchmarks and didn't noticed any performance degradations.
Seems this PR simplifies the implementation of quantile cpu DMatrix :)

[trivialfis]

Yeah, I'm planning categorical data support for CPU, hoping that I can do some cleanup first instead of piling new code.