Improve Integration Tests

[LarsKue]

Our integration tests have so far only been running on Two Moons, which is a niche scenario with no summary network. Ideally, we should have an array of scenarios that we can run integration tests on, in an efficient manner. This would include, for example:

• A scenario with and without a summary network
• Different data shapes
• Online and offline training, as well as training in rounds
• All different kinds of approximator tasks, e.g. Model Comparison, Point Estimation, ...
• Varying sets of metrics and networks
• Running our example Jupyter Notebooks

[LarsKue]

Related to #254

[stefanradev93]

Agreed. We can run these CIs with tiny networks between changes in minor versions.

[LarsKue]

Should be mostly covered as we implement new examples and test them as implemented in #393