Missing value handling

[mnwright]

This is a quick and dirty implementation of missing value handling as done by other packages such as xgboost, lightgbm and scikit-learn. When splitting a node, missing values are ignored for calculating an initial split criterium value (i.e., decrease of impurity). Then for the best split, the missings are tried in the left child node and in the right child node and the choice is made based again on the split criterium value. For prediction, this direction is saved as the "default" direction. If a missing occurs in prediction at a node where there is no default direction, it goes left.

Caveats:

• So far only implemented for classification with the default (Gini) splitting, also not for probability=TRUE
• Only missings in the predictors x are handled, not in the outcome y
• Potential slow implementation, have to check computation speed
• Dirty implementation with repetitive code and no shortcuts if no data is missing at all
• Not well tested

[mnwright]

Now this is also implemented for probability and regression. It's also better tested and faster (no impact if no missing data).