Enhancing Domain Adaptation through Prompt Gradient Alignment Lam Tran