[patrick-immometrica]

Changes proposed in this pull request:

• reduce() on I images adds up to four INT32 pixels as int before the result goes into the double sum, so it overflows once values are above about 2**29.

from PIL import Image

im = Image.new("I", (4, 4), 2**31 - 1)
print(im.reduce(2).getpixel((0, 0)))  # -1
print(im.resize((2, 2), Image.Resampling.BOX).getpixel((0, 0)))  # 2147483647

• Cast the first term to double so the whole expression is evaluated as double.