Evaluation program continues after metrics are calculated

[dahlmolokke]

This happened specifically when evaluating the usps model:
python -m src.evaluation --dataset usps --checkpoint-path weights/usps.pth --device cpu

It continues to run even after printing the results in the CLI

[dahlmolokke]

It happens when num-workers>0. We could either set --num-workers 0 by default, or manage the workers using for instance torch.multiprocessing.set_start_method("spawn", force=True) before the DataLoader is created.
The metrics print fine, so the hang is at shutdown: with num_workers>0 the DataLoader forks workers that can wedge during cleanup and block the main process from exiting. Confirmed that --num-workers 0 fixes it on usps/CPU.