Carl example

[TheEimer]

Adds a CARL generalization example (Kirk protocol). To that, some stuff was also changed in the library:

• option for context set loading in CARL
• additional plotting functions for multi-stage generalization plot
• changes to how contexts are logged (full train/eval sets and then only context id)
• no additional resets in evaluate (since we have an autoreset and this messes with context setting for round robin envs)

This should streamline working with fixed context sets nicely. It doesn't gel well with dynamically sampled contexts, though. My idea is to add a toggle how to log, but I don't know there are maybe smarter ways?