arxiv Simple and Scalable Strategies to Continually Pre-train Large Language Models