Stochastic Second-Order Methods Provably Beat SGD For Gradient-Dominated Functions