Segment First or Comprehend First? Explore the Limit of Unsupervised Word Segmentation with Large Language Models