Debiased Machine Learning without Sample-Splitting for Stable Estimators