Stochastic Subgradient Descent Escapes Active Strict Saddles