Enable more flexible channel map name derivation from files

[d33bs]

@MattsonCam was working with files which look like:

│   ├── 512_768_1024_1280__cell_masks__generated_prediction_00.tiff
│   ├── 512_768_1024_1280__cell_masks__original_00.tiff
│   ├── 512_768_1024_1280__F7-3_TRANS_00.tiff
│   ├── 768_1024_1024_1280__cell_masks__generated_prediction_00.tiff
│   ├── 768_1024_1024_1280__cell_masks__original_00.tiff
│   └── 768_1024_1024_1280__F7-3_TRANS_00.tiff

Needing to use the TRANS part of these filenames to help compose Zarr-based images. This didn't work because image_set_to_arrays looks at the second split value (using _'s) for the key name from the channel_map. We could fix this by making it more flexible to find this information.

We could simulate these files by modifying the existing random sample data and leveraging filenames like these.