DivScene: Benchmarking LVLMs for Object Navigation with Diverse Scenes and Objects