Towards Understanding Cooperative Multi-Agent Q-Learning with Value Factorization