Advancing the Robustness of Large Language Models through Self-Denoised Smoothing