add eps to probs before applying log to avoid infs

[michalwols]

Without it large logits lead to infs and nans due to torch.log(0)

before:

asl = AsymmetricLoss()
asl(torch.Tensor([[-100]]), torch.Tensor([[0]]))
>> tensor(nan)

after:

asl = AsymmetricLoss(eps=0)
asl(torch.Tensor([[-100]]), torch.Tensor([[0]]))
>> tensor(nan)

asl = AsymmetricLoss()
asl(torch.Tensor([[-100]]), torch.Tensor([[0]]))
>> tensor(-0.)

[mrT23]

Thanks @michalwols we didnt encounter these instabilities in our training, but indeed it is important protection.