Fix size calculation of Image.thumbnail()

[orlnub123]

The current size calculation is fine if only one axis is smaller or if the image is wide, but it trips up when both axes are smaller and the image is tall. Here's an example:

from PIL import Image

im = Image.new("L", (145, 100))
im.thumbnail((50, 50))
assert im.size == (50, 34)  # This is fine, the image is wide.

im = Image.new("L", (100, 145))
im.thumbnail((100, 50))
assert im.size == (34, 50)  # Also fine, only one axis is smaller.

im = Image.new("L", (100, 145))
im.thumbnail((50, 50))
assert im.size == (34, 50)  # Oh no! It returned (35, 50)!

Someone made a pull request that fixed this in 2016, but it looks like he closed it shortly after due to a CI failure.

[radarhere]

I imagine that @homm, as the author of #4256, would like to review this?

[homm]

Looks like this happened due to double rounding in both branches.

[orlnub123]

These latest commits have a fix related to the issue previously discussed. Rounding up the size contradicts the documentation which states:

This method modifies the image to contain a thumbnail version of itself, no larger than the given size.

So now when you request a thumbnail with a maximum size of (99.9, 99.9), you get back an image with the size of (99, 99) instead of (100, 100).

As far as I can tell passing float values to size is deprecated, so that commit can be reverted when the switch is made.

[homm]

I have an interesting question. Let's say we have an image 100×30 pixels and requested 75×75 thumbnail. The result will be 75×22.5 which will be rounded to 75×22. It seems that there should be no difference in which direction we're rounding, up or down, both options should be equal.

But the thing is, 100/30 = 3.3(3), 75/22 = 3.4(09) and 75/23 ≈ 3.261 and this is a bit closer to 100/30, but we still have an answer 75×22. Why?

[orlnub123]

I'd like to point out that the current implementation had the same issue. It happens because the rounding is indiscriminate towards the aspect ratio. I've added a commit that fixes the issue, improvements are welcome.

[hugovk]

Thanks!