Grasping the Essentials: Tailoring Large Language Models for Zero-Shot Relation Extraction