When creating a decision tree, does one typically want to limit the number of leaf nodes or the number of samples per leaf?