Dual Debiasing: Remove Stereotypes and Keep Factual Gender for Fair Language Modeling and Translation