Do Large Language Models Solve ARC Visual Analogies Like People Do?