Whispering Experts: Neural Interventions for Toxicity Mitigation in Language Models