Subsample data points at large nodes

[tyler-tomita]

The option/extent of subsampling at large nodes could potentially substantially reduce the amount of sorting while not hurting classification performance.