LLM-Based Multi-Agent Reinforcement Learning: Challenges and Future Directions