Pipeline should use the chain function of its learner

[Larsvanderlaan]

At the moment, the Pipeline object is using the default "chain" which simply chains the predictions of its second learner. However, it would be most natural if Pipeline uses instead the chain function of its second learner. This is in particular needed when you are pipelining two data-processors/screeners.

[Larsvanderlaan]

As a note, most screening/preprocessing learners predict the screened covariate matrix so that this isn't an issue. However, I am working with a custom learner that doesn't do this.