Scaling the Convex Barrier with Sparse Dual Algorithms