Fast Convergence of Regularized Learning in Games