Improving Training Stability for Multitask Ranking Models in Recommender Systems