Dynamic Regret for Strongly Adaptive Methods and Optimality of Online KRR