LMFusion: Adapting Pretrained Language Models for Multimodal Generation