Leveraging Positional Bias of LLM In-Context Learning