Scaling Laws for Generative Mixed-Modal Language Models