IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models