OPUS: Towards Efficient and Principled Data Selection in Large Language Model Pre-training in Every Iteration