Novelty-Guided Data Reuse for Efficient and Diversified Multi-Agent Reinforcement Learning