With a Little Push, NLI Models can Robustly and Efficiently Predict Faithfulness