Bugfix: Resizing masks containing over 65,535 cells not supported by OpenCV

Related to issue #937

When Cellpose detect over 65,535 it casts the masks vector to uint32. However, this format is not natively supported by OpenCV, causing its resize function to crash. See issue #937 for more information.

This PR proposes to use float32 instead, which is supported by OpenCV, and includes the following changes

If the datatype is uint32, first cast to float32, resize using cv2, round and cast resulting matrix to uint32.
Throw a warning if casting to float32 happens
Adding tests to confirm that previous and new function work as expected

This PR also considers three implications:

Time complexity: Casting datatypes is driving the increase in time complexity, causing a 5x to 30x increase. However, in this case it is almost neglectable. A 10,000 x 10,000 image on 32 cores then requires 0.47s compared to 0.02s.
Memory requirements: Float32 will increase memory requirements. The implications are not yet tested.
Numerical stability: Multiple tests have shown that uint16 and the proposed approach which casts to float32 yield identical results.

Reproducible examples

import cv2
import numpy as np

img16 = np.random.randint(0, 65535, size=(1000, 1000, 3)).astype("uint16")
img32 = img16.astype("uint32")

# UINT16: Succeeds
img16r = cv2.resize(img16, (300, 600))

# UINT32: Fails
img32r = cv2.resize(img32, (300, 600))

# FLOAT32: Succeeds (proposed approach)
img32r = cv2.resize(img32.astype("float32"), (300, 600)).round().astype("uint32")

# Identity
(img16r == img32r).mean()

MouseLand / cellpose

Bugfix: Resizing masks containing over 65,535 cells not supported by OpenCV #938

Codecov Report