Image.crop() size depends on position when a box edge is at .5 (round half to even)

[sungbin1015]

I was checking how we turn float boxes from an object detector into crop() calls and noticed that crop() rounds each edge separately with Python's round(). Since Python 3 rounds halves to even, a box of the same width can come out a pixel bigger or smaller depending on where it sits. It's easiest to see with a centre-based box of odd width:

from PIL import Image

im = Image.new("RGB", (200, 100))
for cx in (100, 101, 102, 103):
    box = (cx - 15.5, 10, cx + 15.5, 20)
    print(box, im.crop(box).size)

print(im.crop((1.5, 0, 2.5, 1)).size)
print(im.crop((2.5, 0, 3.5, 1)).size)

(84.5, 10, 115.5, 20) (32, 10)
(85.5, 10, 116.5, 20) (30, 10)
(86.5, 10, 117.5, 20) (32, 10)
(87.5, 10, 118.5, 20) (30, 10)
(0, 1)
(2, 1)

That's on 13.0.0 and current main. A 1px box at 1.5..2.5 gives an empty image, and the same box shifted by one pixel gives 2px. The rounding was added in #1745 (for #1744) while Pillow still supported Python 2, where round() rounded halves away from zero, so I suspect the parity effect came with the move to Python 3 rather than being chosen. It's easy for us to work around by rounding before calling crop(), so nothing is blocked on our side.

Is the half-to-even behaviour intended for crop(), or would you accept a change that rounds halves the same way everywhere, e.g. floor(x + 0.5)?

[radarhere]

A similar issue has been raised about drawing shapes in #10228