Allow subsampling of training data for optimization

[AKuederle]

Sometimes you want to run your optimization only on a subset of the training data (i.e. to speed up the optimization), or just run HyperOpt on some data.

At the moment, this is not easily possible. In particular the case, where you want to do the final training on the full dataset.

We should think about possibilieties to support this.