CityBench: Evaluating the Capabilities of Large Language Model as World Model