Zero-Shot Position Debiasing for Large Language Models