Robust Deep Learning as Optimal Control: Insights and Convergence Guarantees