The Role of Diverse Replay for Generalisation in Reinforcement Learning