Towards Understanding Linear Value Decomposition in Cooperative Multi-Agent Q-Learning