Towards Formal XAI: Formally Approximate Minimal Explanations of Neural Networks