Improving Adversarial Robustness of DEQs with Explicit Regulations Along the Neural Dynamics