OCR-VQGAN: Taming Text-within-Image Generation