[ENH] Second revision of the classifier

[oesteban]

• Set matplotlib plugin to agg in docker image
• Migrate scalings to sklearn pipelining system
• Add Satra's feature selection for RFC
• Make model selection compatible with sklearn Pipeline
• Multiclass classification
• Add feature selection filter based on Sites prediction (requires pinning to development sklearn-0.19)
• New RobustLeavePGroupsOut, which is a leave-one-site-out split that excludes test folds without positive samples. In practical terms: removes the folds for the 2 sites without bad cases in the ABIDE labeling. This allows replacing the buggy RobustGridSearchCV with the standard GridSearchCV of sklearn.
• Choice between RepeatedStratifiedKFold and RobustLeavePGroupsOut in mriqc_clf
• Write cross-validation results to an .npz file.
• Added XGBoost as an alternative for RFC
• Report RFC.oob_score.