Is Incoherence Surprising? Targeted Evaluation of Coherence Prediction from Language Models