Selecting the best model in scikit-learn using cross-validation