Tree and ensemble classification: do not require label metadata

spark.ml DecisionTreeClassifier, RandomForestClassifier, and GBTClassifier require that the labelCol have metadata specifying the number of classes. Instead, if the number of classes is not specified, we should automatically scan the column to identify numClasses.

This differs from SPARK-7126 in that this requires labels to be indexed (but without metadata). This issue is not for supporting String labels.

Note: This could cause problems with very small datasets + cross validation if there are k classes but class index k-1 does not appear in the training data. We should make sure the error thrown helps the user understand the solution, which is probably to use StringIndexer to index the whole dataset's labelCol before doing cross validation.