Add library of tasks for "intelligence tests" to quickly assess model capabilities

[MischaPanch]

The goal is to have an automatic evaluation of model capabilities in the design space without having the developer look at the output.

We can develop the library of (binary) tests in tandem with e.g. #89.

Some things we might test:
Pixel-perfection of refactoring comes to mind, but other things are possible. Could be executed on demand with pytest (marking as skipped based on a condition), or in a dedicated script