Exploring Imbalanced Annotations for Effective In-Context Learning