[EM] Optimization for deep trees.

[trivialfis]

• Decouple the row partition batch size from the driver batch size. This will allow us to process more nodes for each data batch.
• Pick a heuristic to use ATS instead of data copy to handle cases where we have a large number of small nodes.
• Make sure a new page that happens to be the last is placed on the host.

The optimization is specifically for GPU external memory. We need some kernel fusion to improve in-core training further.

todo:

• Perf test
• Make sure the benchmark is deterministic.

in-core

depth/branch	PR	master
6	160.2554975748062	160.7703891992569
12	275.5357859134674	277.49394392967224

out-of-core

	PR	Master
12	2990.530539	7855.628397

[trivialfis]

cc @rongou .