Fix group-wise evaluation

[naga-karthik]

currently, the wrapper script compute_metrics_reloaded.py averages the results over all cases (i.e. when prediction and GT are empty, prediction empty GT not empty, etc. etc.).

When pred and GT are both empty, the DSC is set to 1 automatically (which is correct as the model has rightly learned to not output a false positive). BUT, a lot of these outputs, skews the DSC in such a way that we don't how the model performs in case where there is lesion (i.e. does it predict the whole lesion, does it predict only partially, etc.)

SO, for this, we want to separate the evalutation of results into two cases: (1) when GT is not Empty (and then average the resutls), (2) when the GT is empty, maybe compute the False Postive Rate.

credit to Julian McGinnis who started this discussion!