Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning