A Multi-Agent Approach for Adaptive Finger Cooperation in Learning-based In-Hand Manipulation