Revisiting Logistic-softmax Likelihood in Bayesian Meta-learning for Few-shot Classification