Extractive Fact Decomposition for Interpretable Natural Language Inference in one Forward Pass