Learn to Follow: Decentralized Lifelong Multi-agent Pathfinding via Planning and Learning