Learning to Collaborate in Multi-Module Recommendation via Multi-Agent Reinforcement Learning without Communication