Implement CIC agent

[milosen]

Strategy:

• create env-wrapper that changes the observation-type to a dict that also has a "skill"-key and generate skill on the environment level; also remove the reward
• change the policies to intercept the flow between the feature extractor and Actor; it should split the observation into skill and rest pass both separatly through their respective encoder and combine the state and skill embeddings to return
• the critic should ignore the skill variable