Confident Sinkhorn Allocation for Pseudo-Labeling