Exploring Architectural Ingredients of Adversarially Robust Deep Neural Networks