Reaching Consensus in Cooperative Multi-Agent Reinforcement Learning with Goal Imagination