Handling of missings (in train + predict)

[pat-s]

Discussed several times in mlr:

• Discuss the changes in #2099 mlr#2204
• Can't predict on randomForest when test set contains NA's in features mlr#1515

AFAICS the current behavior in mlr is the one of this PR: mlr-org/mlr#2099

[mllg]

Missings are now handled, with the following policy:

• If a learner is capable of handling missing values during train(), it should get the missings property.
• Learners which cannot handle missing values in the test set should predict NA for these observations.
• Predicting NA results in an exception unless you have a fallback learner defined. All rows with NA observations are imputed with the predictions of the fallback learner.

I know that this is not perfect, and that there might be some rare occasions where you need more flexibility. However, I believe that this is a statistically sound approach (unlike na.rm = TRUE stuff during performance assessment). Additionally, you can always impute missing values with a PipeOp.