Long-Text-to-Image Generation via Compositional Prompt Decomposition