add check for missings in newdata

[QuayAu]

Fix for #1515

[mb706]

Usually the prediction made for one row of data should not depend on predictions made for other data rows, or be even dependent on the presence or absence of other data rows. I would suggest to extend this logic here: Don't throw an error if NAs are present, and instead predict NAs for the rows that contain NAs.

In practice, if a Learner does not support NAs:

1. Remove data rows that contain NAs from input data
2. Perform prediction on this reduced data set
3. Insert NAs in the predictions returned by the Learner at the position of rows that were remove in 1.

[pat-s]

Superseeded by #2099