Prepare children histograms, one per column.
Typically, histograms are created with a level-dependent binning strategy.
For the histogram of the current split decision, the children histograms are left/right range-adjusted.
Any histgoram can null if there is no point in splitting
further (such as there's fewer than min_row elements, or zero
error in the response column). Return an array of DHistograms (one
per column), which are bounded by the split bin-limits. If the column
has constant data, or was not being tracked by a prior DHistogram
(for being constant data from a prior split), then that column will be
null in the returned array.
currentHistos - Histograms for all applicable columns computed for the previous split finding process
way - 0 (left) or 1 (right)
splat - Split point for previous split (if applicable)
parms - user-given parameters (will use nbins, min_rows, etc.)
Array of histograms to be used for the next level of split finding