fix: remove redundant final_activation layer in model.py

[JCasaraconn]

Remove redundant final_activation layer in the forward() method of AbstractUNet.

The layer is redundant because it is only applied during validation, however the loss function of segmentation models already applies a final activation layer. Therefore the final_activation layer is applied twice in these instances.

[JCasaraconn]

I'm still thinking whether or not this means the validation eval should apply the final_activation layer, similar to the train eval as described here:

pytorch-3dunet/pytorch3dunet/unet3d/trainer.py

Lines 213 to 222 in 5917023

	if isinstance(self.model, nn.DataParallel):
	final_activation = self.model.module.final_activation
	else:
	final_activation = self.model.final_activation
	if final_activation is not None:
	act_output = final_activation(output)
	else:
	act_output = output
	eval_score = self.eval_criterion(act_output, target)

I'm curious, what are your thoughts?