[EM] Merge GPU partitioning with histogram building.

[trivialfis]

• Stop concatenating pages if there's no subsampling.
• Use a single iteration for histogram build and partitioning. It's critical to reduce the amount of looping over the cache.

This PR does not include some optimization work, as I want to redo some of the analyses before upstreaming optimizations.

After this PR, the basic framework for GPU batch-based training is considered complete. We will then focus on feature integration and performance optimization.

Todos:

• Tests
• Perf check.

Known limitations:

• Changing sampling during training can lead to issues. We will have a switch to decide whether page concatenation is needed. For now, sampling turns on concatenation automatically.

[trivialfis]

cc @rongou .