Enhancing Robustness of LLM-Driven Multi-Agent Systems through Randomized Smoothing