Optimization for deep trees on distributed system.

Currently, XGBoost's performance is sub-optimal when growing deep trees (max_depth > 16) on distributed systems, partly due to memory allocation and partly due to allreduce call. The CPU utilization is relatively low (<60%).

Start by profiling distributed training with max_depth greater than 16, focusing on memory allocation, allreduce calls, and CPU utilization. The issue names no files, tests, or entry points, so first identify the relevant deep-tree training and distributed communication paths. Done means a measured improvement in the reported performance bottlenecks.

Written by the indexing model from the issue text.