CoAnnotating: Uncertainty-Guided Work Allocation between Human and Large Language Models for Data Annotation