Aligning Instruction Tasks Unlocks Large Language Models as Zero-Shot Relation Extractors