Create bounding-box ROI from label images

[jluethi]

New task: Take a label image and create new ROIs based on the bounding boxes of objects. Useful e.g. for organoids after organoid segmentation in 2D.

See #45 for issues we'll need to handle regarding overlap

[tcompa]

Preliminary plan, with @mfranzon:

1. Consider example input like this one.
2. Write a function that takes a label array (2D/3D array of integers) and extract all (2D/3D) bounding boxes.
3. Use this function in another function that simply writes new bounding-box ROIs to a AnnData table in the zarr array. This first version does not perform any check on overlaps.
4. Include this function at the end of cellpose_segmentation (first) and napari_workflow_wrapper (later one, once it's polished).
5. Add a check for overlapping bounding-box ROIs. By now we should just raise a warning - see Establish ROIs for organoids #45 for more about this.

[jluethi]

I like this plan. Step 4 should be optional, because I'm not sure if we want to have ROI tables for all segmentation outputs and we certainly may have napari workflows where we don't need the ROI tables.
Because ROI tables are light-weight, we can default this to True, create ROI tables by default. Will be interesting to explore whether ROI tables are also useful for single cells (many more than organoids) and whether the whole architecture scales well for having 1000s of ROIs. If not, we can turn if off by default.

Regarding 1) and ignoring overlap for the moment: Good start. But if we want to test processing by ROI afterwards, we'll have to look at an example where there is no overlap. We may have one from Max from the Zebrafish that I'm currently testing, will report back if it's suitable. But we anyway need to ROIs calculated & parsed first :)

[tcompa]

Write a function that takes a label array (2D/3D array of integers) and extract all (2D/3D) bounding boxes.

By now we don't need to support 2D arrays, since both cellpose_segmentation and napari_workflow_wrapper return their 2D label output as 3D arrays (in cellpose_segmentation this is always true, in napari_workflow_wrapper this takes place at a certain point during the output handling).

If it is needed for some use case, we can always include an explicit handling of 2D arrays later.

[tcompa]

Closed by #192.

Re: overlapping bounding boxes, at the moment we simply raise a ValueError. The future steps are part of:

• Store masked array to zarr #194
• Associate optional mask to each ROI (in tables), and use it for I/O selection #195
• Establish ROIs for organoids #45