On the Geometry of Regularization in Adversarial Training: High-Dimensional Asymptotics and Generalization Bounds