Safe and Robust Multi-Agent Reinforcement Learning for Connected Autonomous Vehicles under State Perturbations