Topological Uncertainty: Monitoring trained neural networks through persistence of activation graphs