Multi-Agent Reinforcement Learning for Deadlock Handling among Autonomous Mobile Robots