PSL: Rethinking and Improving Softmax Loss from Pairwise Perspective for Recommendation