Enhancing Robustness of Offline Reinforcement Learning Under Data Corruption via Sharpness-Aware Minimization