Make torch.nn.functional.multilabel_soft_margin_loss more stable

[vadimkantorov]

Currently it calls sigmoid and then binary_cross_entropy (which does log internally):
https://github.com/pytorch/pytorch/blob/master/torch/nn/functional.py#L1796-L1797

Is it not possible to compute it as: torch.mean(-(y * F.logsigmoid(x) + (1 - y) * F.logsigmoid(-x))) ?

y * log(1 / (1 + exp(-x)) + (1 - y) * log(1 - 1 / (1 + exp(-x))) =
y * logsigmoid(x) + (1 - y) * log((1 + exp(-x) - 1) / (1 + exp(-x))) = 
y * logsigmoid(x) + (1 - y) * log(1 / (1 + exp(x))) = 
y * logsigmoid(x) + (1 - y) * logsigmoid(-x)