Rethinking the Optimization of Average Precision: Only Penalizing Negative Instances before Positive Ones Is Enough