Improve CPU performance

[RAMitchell]

Bring legate-boost cpu performance somewhat in line with xgboost/lightgbm.

[RAMitchell]

#221 has dramatically improved CPU performance but we still have a gap of around 3x compared to XGBoost

Benchmarking attributes this mostly to the binning process which occurs for legate-boost in every tree but only once in XGBoost. To improve this we would need to either rearchitect legate-boost to do the binning once (and lose the 'self-contained' implementation of models) or simply make the binning process faster.

Another issue is the inefficient CPU allreduce implementation (#72), which hurts our strong scaling performance at 32 CPUs.

[magnatelee]

@RAMitchell I suspect the numbers are in the regime where the runtime overhead isn't a bottleneck, otherwise please let the legate team know about the issue.