Remove kwargs in learn and update

[MischaPanch]

That's a tricky one since learn implementation are currently not cleanly separated (violate Liskov substitution) and rely on kwargs in order to remain functional. E.g. some implementations of learn take the repeat kwarg (the on-policy ones), and some don't.

Generally, the control flow between trainer.update_policy and policy.learn may need to be redesigned or at least improved

[MischaPanch]

@maxhuettenrauch @opcode81

This is the issue we talked about.

Solving it will likely be done by introducing a new Algorithm abstraction that incorporates some functionality of BasePolicy and some of BaseTrainer. This is a large change, but it's probably the only way to adequately address this issue.

Feel free to extend the issue description with more details of the current plan for solving this.