Distributed Circle Formation Control for Quadrotors Based on Multi-agent Deep Reinforcement Learning