Tune criterion for RF / XT

[dholzmueller]

For classification with RF/XT, we could optimize the criterion in ['gini', 'entropy']. In my tests, one could just always use 'entropy' if logloss is the target metric, but I suspect that 'gini' might be better for other target metrics.
Since the criterion needs to be different for classification and regression, one could add a hyperparameter classification_criterion which is tuned in ['gini', 'entropy'] and ignored for regression. This might require changing AutoGluon code, though.