Do Vision-Language Models Have Internal World Models? Towards an Atomic Evaluation