Skill Learning via Policy Diversity Yields Identifiable Representations for Reinforcement Learning