Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoor Attacks