Adaptation for instance segmentation inference

[titusgriebel]

Changes to the original CellViT:

• torch-em dataloaders implemented --> .tiff input for labels and images
• some pre-processing for the torch-em loader output implemented for downstream compatibility
• .tiff-writing of the sole instance segmentation masks (without gt, binary, raw as provided by the default script) to a separate directory implemented