predict_time can be (kind of) wrong

[sebffischer]

When calling learner_predict(), the predict_time in the state is overwritten.
When doing a resampling and making predictions for different sets (in my case the holdout set and the test set), the time_predict is not (as I expected) the total prediction time, but only the prediction time for one of the predictions.

It would be great to have access to the different prediction times.

[mllg]

Store cumulative time for all predict sets + document.