DivTOD: Unleashing the Power of LLMs for Diversifying Task-Oriented Dialogue Representations