Add cellpose training and evaluation

[hermancollin]

This PR adds cellpose to the benchmark. I used the cyto3 model weights as a starting point, and trained for 500 epochs on all datasets.

For the inference procedure, I used the following flags:

cellpose --pretrained_model PATH_TO_MODEL --dir PATH_TO_DATA --img_filter _SUFFIX --mask_filter _seg-cellpose --verbose --use_gpu --save_png --no_npy

Additionally, for the TEM1 and TEM2 datasets, I had to use the --diameter option and set it to the model diam_labels ("mean diameter of training ROIs") .

Then, the evaluation script fetches the cellpose predictions and computes object-detection metrics for it with the --eval_cellpose option.