Unveiling the Capabilities of Large Language Models in Detecting Offensive Language with Annotation Disagreement