Tight Robustness Certificates and Wasserstein Distributional Attacks for Deep Neural Networks