kknn should give an informative error message when missing values lead to an error

[mb706]

> l <- lrn("classif.kknn")$train(tsk("pima"))
> l$predict(tsk("pima"))
Error in check_prediction_data.PredictionDataClassif(pdata) : 
  Assertion on 'as_factor(pdata$response, levels = lvls)' failed: Must have length 768, but has length 392.

Pretty sure this is because of missing values (kknn just ignores them and creates a prediction for the rows that don't contain missing values), the error message should make this clear.

Maybe this is a good time to lobby for splitting up the 'missing' property into train-missing and predict-missing or smth like that.

[mllg]

We now throw an exception during train.