Double Variance Reduction: A Smoothing Trick for Composite Optimization Problems without First-Order Gradient