Batch Size or Negatives? A Selection Rule for Memory-Constrained Recommender Training