learner_cv use cv-models for inference

[mb706]

think about: should this give out a multiplicity during prediction? Then during training, it could give just N x what it already returns, during inference the multiplicity of predictions of all models. Or we do averaging in the learnercv (less flexible?). I think we currently don't have any POs that would be compatible with option 1.

[mb706]

maybe we want to do multiplicity output during training and prediction, and have something that averages in both cases

[mb706]

this way we could have bagging instead of CV, or repCV, as well