Provably Robust and Plausible Counterfactual Explanations for Neural Networks via Robust Optimisation