From One to Many: Expanding the Scope of Toxicity Mitigation in Language Models