Words Worth a Thousand Pictures: Measuring and Understanding Perceptual Variability in Text-to-Image Generation