I am planning on submitting several PRs in an attempt to merge #5041 in slowly, with the ultimate goal being a clean implementation of multithreaded decision tree building so that Gradient Boosting can be faster. With one of the main concepts merged (#5203), here is a list of separate PRs which I'd like to merge in the near future.

• Reorganize _tree.pyx into several files (see PR [MRG+1] split tree module into several packages #5230)[merged]
• Add proxy impurity improvement methods to both Gini and entropy (see PR [MRG] Proxy improvement methods added to entropy/gini #5233)[closed]
• Reevaluate constant feature caching [closed]
• Support sparse data for gradient boosting (see PR [MRG+1] Merge PresortBestSplitter and BestSplitter #5252)
• Add caching of computation between different split levels to avoid recomputation
• Ensure feature importance converge in ensemble (see PR [MRG+1] Stronger tests for variable importances #5261)
• Add tests to ensure the correctness of impurity values, wrt hand-computed values on toy data.

Longer range goals which I'd like to work towards (but have no clear plan as of right now) are the following:

• Add an approximate splitter
• Add multithreading support for single decision trees
• Add a partial fit method for tree building
• Support categorical variables
• Support missing values
  Issue [RFC] Missing values in RandomForest #5870, PR ENH Adds missing value support for trees #23595, FEA Add support for missing values in tree estimators with criterion="absolute_error" by greatly simplifying the logic #32119

At this point, it will be clearer to me what specific changes to Splitter, Criteria, and TreeBuilder need to be added to make multithreading a possibility. @glouppe @arjoly @GaelVaroquaux @pprett if you have any comments, I'd love to hear them.