Unveiling A Core Linguistic Region in Large Language Models