On the Computational Efficiency of Adaptive and Dynamic Regret Minimization