BinaryLogisticRegressionSummary, GLM Summary should handle non-Double numeric types

Say you have a DataFrame with a label column of Integer type. You can fit a LogisticRegresionModel since LR handles casting to DoubleType internally.

However, if you call evaluate() on it, then this line does not handle casting properly, so you get a runtime error (MatchError) for an invalid schema: https://github.com/apache/spark/blob/2cd327ef5e4c3f6b8468ebb2352479a1686b7888/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala#L863

We should handle casting. And test evaluate() with other numeric types.

*ALSO* We should check elsewhere in logreg and other algorithms to see if we can catch the same issue elsewhere.