Stochastic Gradient Descent for Nonconvex Learning without Bounded Gradient Assumptions