Mitigating Text Toxicity with Counterfactual Generation