Getting SDCA multiclass weights requires using a MulticlassLogisticRegressionPredictor, which is strange

[GalOshri]

As shown here, getting the predictor weights from multiclass-SDCA requires using MulticlassLogisticRegressionPredictor. This is counter-intuitive as I would expect that predictor to be used just for Logistic Regression. Perhaps something more general like MulticlassLinearPredictor should be created/used?

[wschin]

Looks like MulticlassLogisticRegressionPredictor should be split into two classes, MulticlassLogisticRegressionPredictor and MulticlassLinearPredictor. SDCA always producing MLRP is obvious wrong because multi-class support vector machine is not multi-class logistic regression. cc @TomFinley.