Breaking the Gradient Barrier: Unveiling Large Language Models for Strategic Classification