How Humans Help LLMs: Assessing and Incentivizing Human Preference Annotators