Add 2D map of summary space as diagnostic to assess model misspecification

[vpratz]

When comparing the summary space two different distributions (e.g., a training distribution and a observed distribution), a two-dimensional map of the high-dimensional summary space can provide first insights. The most common methods for this are probably UMAP and t-SNE.

The easiest way to obtain those would probably be scikit-learn, which we just removed as a dependency.
@stefanradev93 @LarsKue What do you think? I consider scikit-learn a relatively lightweight dependency, which offers many convenience functions. Is it a priority to go without it, or would it be ok to include it again to offer the functionality described above?

[LarsKue]

I have no strong opinion towards the scikit-learn dependency. I believe @stefanradev93 removed it because we were only barely depending on it.