Improving Input-label Mapping with Demonstration Replay for In-context Learning