Is Sarcasm Detection A Step-by-Step Reasoning Process in Large Language Models?