Beyond Worst-case Sequential Prediction: Adversarial Robustness via Abstention