Enhancing Certifiable Semantic Robustness via Robust Pruning of Deep Neural Networks