The SCELoss may not be correctly computed

[Cascol-Chen]

For example, when the input is [0.9, 0.1], the label is [0] and n_bins=1, the SCELoss will result in a great value because the acc_matrix following code's calculation is [True, True] (treating 0.1 as right answer), and $conf_0=0.9, acc_0=1, conf_1=0.1, acc_1=1$. Since $conf_1$ and $acc_1$ differs greatly, the loss will be huge.