Adversarial Training in Low-Label Regimes with Margin-Based Interpolation