RegressionEvaluator return 0 for MSE && RMSE if all scores are NaN

It seems that RegressionEvaluator filters out all `NaN` scores when calculating metrics. So if a predictor produces `NaN` only for the entire evaluation dataset, the RMSE and MSE from RegressionEvaluator would be 0, which is ironically the best result it can get.

While it may need further discussion, I wonder if it's necessary to _FILTER_ out the `NaN` value when calculating metrics, as `NaN` in prediction result usually indicates a model is not well-trained, and thus should be reflected in evaluation metrics.