From Large Language Models to Small Logic Programs: Building Global Explanations from Disagreeing Local Post-hoc Explainers