Demonstration Augmentation for Zero-shot In-context Learning