Extract evaluate splits from CPU hist.

[trivialfis]

Other than modularizing the split evaluation function, this PR also removes some more functions including InitNewNodes and BuildNodeStats among some other unused variables. Also, scattered code like setting leaf weights is grouped into the split evaluator and NodeEntry is simplified and made private. Another subtle difference with the original implementation is that the modified code doesn't call tree[nidx].Parent() to traversal upward.

perf

I think it's faster after the PR. URL is only run for 100 rounds, others are 500 rounds.

	HIGGS	covtype	URL
master	180.24662968399934	76.94775568600744	196.1527223587036
PR	177.9364239120041	68.41261254699202	180.37877798080444

[trivialfis]

cc @ShvetsKS @SmirnovEgorRu

The unused_rows_ is not used in the builder, but used in the test. I think it's legacy?

[ShvetsKS]

cc @ShvetsKS @SmirnovEgorRu

The unused_rows_ is not used in the builder, but used in the test. I think it's legacy?

yes. It's not used now, gradients are set to zero instead.

[trivialfis]

@ShvetsKS Thanks for the reply. Could you help review this PR? I will remove the unused_rows_ in upcoming refactors as I'm migrating the approx implementation to the fast hist code base.

[SmirnovEgorRu]

LGTM

[hcho3]

LGTM. Two comments:

• It's nice that the CPU Hist is becoming more like GPU Hist. Consistency makes the whole codebase more legible.
• "Evaluator" is a nice abstraction. I like how we no longer need a mock for testing split evaluation.

[ShvetsKS]

LGTM