New Insight into Hybrid Stochastic Gradient Descent: Beyond With-Replacement Sampling and Convexity