Rethinking Offensive Text Detection as a Multi-Hop Reasoning Problem