Use heuristic to select histogram node, avoid rabit call

[RAMitchell]

This PR considerably improves performance for the distributed GPU algorithm. We use the sum of Hessian in the left and right node to estimate which has fewer training instances instead of using rabit to sync and calculate this exactly. In particular it improves latency.

Performance numbers from training on a DGX-1 with https://github.com/NVIDIA/gbm-bench

dataset	old	new	speedup
airline	68.4073	59.2706	1.15
bosch	20.12437	16.1119	1.25
covtype	81.049	59.8792	1.35
epsilon	306.2088	311.6957	0.98
fraud	8.3944	6.7949	1.24
higgs	34.4294	24.0335	1.43
year	25.8719	15.3342	1.69

I also removed the old HostAllReduce function, which is redundant now that we no longer manage multiple GPUs with threads.