Describir: Property Checking with Interpretable Error Characterization for Recurrent Neural Networks