Convex Dual Theory Analysis of Two-Layer Convolutional Neural Networks with Soft-Thresholding