Sharpness-Aware Training for Accurate Inference on Noisy DNN Accelerators