[BUG] region_props handling of NumPy array

[JHancox]

Describe the bug
In cucim.skimage.measure, region_props allows you to pass in a numpy object and some properties work okay (e.g. bbox) whereas others produce an error when evaluated (e.g. centroid). This inconsistency is somewhat confusing. It seems that ndarray is only partially supported internally.

Steps/Code to reproduce bug

import cupy as cp
from cucim.skimage import measure

a = np.array([[0,0,0,0,0,0,0],
              [0,1,1,1,0,0,0],
              [0,1,1,1,0,0,0],
              [0,1,1,1,0,0,0],
              [0,0,0,2,2,2,0],
              [0,0,0,2,2,0,0],
              [0,0,0,3,3,0,0]])

# Uncomment and it works. Omit this and some properties work 
# a = cp.array(a)

res = measure.regionprops(a)

for r in res:
    print(r.bbox, r.centroid)

Expected behavior
All properties should either work without error or should fail consistently

Environment details (please complete the following information):
Running Vista2D container, which includes cucim v23.12.1

Additional context

[grlee77]

Thanks for reporting the issue, @JHancox. I will work on a fix for this.

FYI, I am currently working on improving regionprops for GPU as I think it is a popular part of scikit-image, but is currently not accelerated well in cuCIM. I am not sure we can improve the case involving calling RegionProps on individual labels as in the code above, but my goal is to provide >10x improvement for properties computed via regionprops_table.

So for GPU use, once the revised implementation is available it should be much faster to do

props = measure.regionprops_table(a, properties=['bbox', 'centroid'])

than

res = measure.regionprops(a)

for r in res:
    print(r.bbox, r.centroid)

The later is problematic as individual regions are often small, so much of the GPU sits idle and there is separate kernel launch overhead for each region. We can avoid both of these issues when computing properties for all labels at once.