Users as Annotators: LLM Preference Learning from Comparison Mode