Memory leak when returning class_specific_scores

[sabeenlohawala]

https://github.com/sabeenlohawala/tissue_labeling/blob/45f1e1ee14491246cbb3c0cd6b17932ef5f91814/TissueLabeling/metrics/metrics.py#L86C9-L86C39

Noticed that GPU memory increases with every batch when class_specific_scores is True and training on the new_kwyk_data. This is not the case when class_specific_scores is False → identify the memory leak