Muffin or Chihuahua? Challenging Large Vision-Language Models with Multipanel VQA