Revisiting Adversarial Training for ImageNet: Architectures, Training and Generalization across Threat Models