All Should Be Equal in the Eyes of Language Models: Counterfactually Aware Fair Text Generation