Logical Reasoning with Span-Level Predictions for Interpretable and Robust NLI Models