Leveraging Approximate Symbolic Models for Reinforcement Learning via Skill Diversity