Fast, Differentiable and Sparse Top-k: a Convex Analysis Perspective