added final activation before eval score calculation

[jens-maus]

This PR applies the same final activation function before calculating the eval score than before calculating the validation score accordingly. This should make eval score calculation consistent with the validation score calculation (e.g. in case the Dice coefficient is used) so that both metrics are more comparable to each other.

[wolny]

Hi @jens-maus, that's a really nice catch! Currently the final_activation is only applied when the model is in the validation mode (see: https://github.com/wolny/pytorch-3dunet/blob/master/pytorch3dunet/unet3d/model.py#L97). This PR correctly applies the final_activation when we want to compute the evaluation metric in the training mode. Thanks!