Enhancing Post-Hoc Attributions in Long Document Comprehension via Coarse Grained Answer Decomposition