ParaPO: Aligning Language Models to Reduce Verbatim Reproduction of Pre-training Data