PNG palette quantized save functionality different between 8.1.0 and 8.3.1

[paulds8]

What did you do?

Create a palette, quantize an image using this palette and save it.

What did you expect to happen?

The palette information is stored correctly. I.e. the list of RGB values is the same in the saved image.
E.g. [0, 0, 0, 209, 254, 76, 217, 3, 3, 10, 66, 192,...]

What actually happened?

The list of RGB values is corrupted.
E.g. [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3,...]

What are your OS, Python and Pillow versions?

• OS: Tested on Windows and Azure Databricks (Ubuntu)
• Python: 3.6 & 3.8
• Pillow: 8.3.1 has the issue

from PIL import Image
from itertools import cycle
import numpy as np

palette = [0, 0, 0, 209, 254, 76, 217, 3, 3, 10, 66, 192]
# palette must contain 768 values (256 * 3)
# repeat the colours until we get there
color_iter = cycle(palette)
expanded_palette_format = []
while True:
    expanded_palette_format.append(next(color_iter))
    if len(expanded_palette_format) == 768:
        break

assert len(expanded_palette_format) == 768
palette_image = Image.new('P', (16, 16))
palette_image.putpalette(expanded_palette_format)

# a new image that we want to apply the palette to
im_to_palette = np.array(
    [
        [[0, 0, 0,],[209, 254, 76], [217, 3, 3], [10, 66, 192]],
        [[0, 0, 0,],[209, 254, 76], [217, 3, 3], [10, 66, 192]],
        [[0, 0, 0,],[209, 254, 76], [217, 3, 3], [10, 66, 192]],
        [[0, 0, 0,],[209, 254, 76], [217, 3, 3], [10, 66, 192]]
    ],
    np.uint8
    )
im = Image.fromarray(im_to_palette).quantize(palette=palette_image, dither=0)

im.save(r"test.png")
saved_im = Image.open(r"test.png")

assert im.getpalette() == saved_im.getpalette()
# AssertionError

Please note that if I specifically pass the 'GIF' argument on save, it has the expected behaviour.

im.save(r"test.png", "GIF")
saved_im = Image.open(r"test.png")

assert im.getpalette() == saved_im.getpalette()

[radarhere]

I've created PR #5696 to resolve this.