Aligning Large Multimodal Models with Factually Augmented RLHF