Refactor recommender signature

[AdrianSosic]

This PR implements a breaking change by refactoring the recommender signature such that:

• it accepts an optional Objective
• it expects training data as a single dataframe in experimental representation

Apart from fixing certain responsibilities (e.g. Campaign now only acts as a meta-data handler, just like it is supposed to) this enables/prepares several new features:

• Users can directly use Recommenders as entry point instead of being forced to go via Campaign
• Because recommenders now operate on experimental representations, they can offer the same interface back to users, e.g. when exposing model-internal things such as acquisition functions and surrogate models
• Additional pre-processing steps that require access to the experimental representation (such as data-augmentation) can now happen inside recommenders

[AdrianSosic]

Hi @AVHopp, this PR might come somewhat surprising to you because it is contrary to what we have agreed on, but there have been new developments that require us to prioritize this. Can give you an update in a call 🙃

[AVHopp]

I tried my best to wrap my head around this :D See all of my individual comments.

[AdrianSosic]

Hi @AVHopp, @Scienfitz: have refactored the code quite a bit during rebasing in order to 1) incorporate your comments and 2) improve the design a bit further. This makes many of the existing threads obsolete since the corresponding code lines no longer exist. Will close them and mark with obsolete after refactoring.

That means basically: let's have another fresh look at the PR 🙃 Mypy problems are now also resolved. There are two remaining open points to discuss for which I'll open separate threads