Fix batch recommendation

[AdrianSosic]

When the non-GP surrogates where added, they were not designed with batch recommendations in mind. Now that batch recommendations can be generally requested, surrogate models providing only marginal posterior information yield unusable results for batch sizes larger than one.

This PR fixes the issue by:

• Making the RandomForestSurrogate produce a proper EnsemblePosterior that is capable of expressing covariance structure, which is the basic requirement for batch prediction
• Disallowing batch sizes larger than one for surrogates that only provide marginal posteriors

[AVHopp]

Not sure whether I fully understand some of the changes made here. In general this looks good, but I'd appreciate some details since I do not know if I understood everything correctly

[AVHopp]

Since all of my question are just minor/my understanding, feel free to resolve as you see fit :)