Stochastic Subgradient Method Converges on Tame Functions