Model-Based Multi-Agent RL in Zero-Sum Markov Games with Near-Optimal Sample Complexity