AC2C: Adaptively Controlled Two-Hop Communication for Multi-Agent Reinforcement Learning