SK-VQA: Synthetic Knowledge Generation at Scale for Training Context-Augmented Multimodal LLMs