Revisiting Logistic-softmax Likelihood in Bayesian Meta-Learning for Few-Shot Classification