Improving Fast Adversarial Training with Prior-Guided Knowledge