Avoid redundant computation in GMM.transform && GLR.transform

In SPARK-27944, the computation for output columns with empty name is skipped.

Now, I find that we can furthermore optimize

1, GMM.transform by directly obtaining the prediction(double) from its probabilty prediction(vector), like what ProbabilisticClassificationModel and ClassificationModel do.

2, GLR.transform by obtaining the prediction(double) from its link prediction(double)