MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents