Cautiously-Optimistic Knowledge Sharing for Cooperative Multi-Agent Reinforcement Learning