Reusing Computation in Text-to-Image Diffusion for Efficient Image Generation