Overhaul of torchseg.losses

[notprime]

Overhaul of torchseg.losses, starting from DiceLoss.

General changes applied to DiceLoss (to be applied later to all other losses):

• switched from assert to warnings and raise;
• eliminated the need of specifying binary, multiclass and multilabels modes when instantiating a loss function, by changing masks shape: now the shape must be [B, C, H, W] for multiclass and multilabels cases, [B, 1, H, W] for binary case. [B, 1, H, W] can be accepted also for multiclass case as long as mask_to_one_hot = True;
• introduction of proper reduction techniques (MEAN, SUM, NONE) and definition of a LossReduction class to easily save the techniques;

Specific changes applied to DiceLoss:

• added the argument power to be applied to the denominator of the DiceCoefficient (still need to add this to the documentation of the class)

[notprime]

@isaaccorley let me know what do you think about the "new" DiceLoss, it seems cleaner and more efficient without the burden of specyfying binary, multilabel and multiclass modes.

I'm not particularly sure about line 135 and 136 in the forward of DiceLoss, it does not make much sense to me for the two reasons I've written in the comment.

I think that we should just rewrite the losses before implementing the new ones