Output soft lesion segmentation

[valosekj]

SCIseg currently outputs binary lesion segmentation. This might not be always ideal. For example, the tissue bridges are usually computed on soft lesion segmentations (manually created using JIM). @jcohenadad thus suggested extending SCIseg inference to output the soft segmentation using dropout and Monte Carlo. After some investigation with @naga-karthik, this could be done by commenting out this line in the nnunet source code (which disable Dropout by default whenever .eval() is run) and modifying the following lines in the plans.json file:

"dropout_op": torch.nn.Dropout,
"dropout _op_kwargs": 0.5,

EDIT: clarified why that line has to be commented out