Safe Multi-Agent Reinforcement Learning for Formation Control without Individual Reference Targets