Stochastic Gradient Succeeds for Bandits