Improving Long-Text Alignment for Text-to-Image Diffusion Models