Localizing Factual Inconsistencies in Attributable Text Generation