Multi-Agent Reinforcement Learning with Control-Theoretic Safety Guarantees for Dynamic Network Bridging