Understanding Partial Multi-label Learning via Mutual Information