Composition and Deformance: Measuring Imageability with a Text-to-Image Model