Improving the Robustness of Quantized Deep Neural Networks to White-Box Attacks Using Stochastic Quantization and Information-Theoretic Ensemble Training