The Pessimistic Limits and Possibilities of Margin-based Losses in Semi-supervised Learning