Efficient Reinforcement Learning for Zero-Shot Coordination in Evolving Games