Accelerating Recommender Model Training by Dynamically Skipping Stale Embeddings