Relaxed child images check to allow for libjpeg

[radarhere]

Reported in conda-forge/pillow-feedstock#128 (comment)

test_get_child_images() in test_file_jpeg.py is failing with libjpeg. This PR relaxes the check to test for a similarity of 2.1, instead of equality. I have confirmed this in pillow-wheels with libjpeg 9e.

[h-vetinari]

Sorry, this was a false alarm!

I got caught out by the fact that test_strip_ycbcr_jpeg_2x2_sampling started re-failing for 9.4 when I switched to libjpeg-turbo, and actually eps=1.2 is fine for test_get_child_images!

Sorry for the mix-up!

Perhaps it would be possible to widen the skip of test_strip_ycbcr_jpeg_2x2_sampling to take into account newer libjpeg_turbo? Currently it only skips 2.0, but we're on 2.1 (and ABI-wise, we're only pinning to 2, without the minor version).

[h-vetinari]

In other words, this PR is still necessary (or at least useful for us), but should be good enough with the following:

[h-vetinari]

Suggested change

	assert_image_similar_tofile(ims[0], "Tests/images/flower_thumbnail.png", 2.1)
	assert_image_similar_tofile(ims[0], "Tests/images/flower_thumbnail.png", 1.2)

[radarhere]

When I test with libjpeg 9e, I also found a difference of 2.0033 - https://github.com/radarhere/pillow-wheels/actions/runs/3821343292/jobs/6500396156

If we're going to change this to improve support, might as well change it to support more libjpeg environments in general.

[radarhere]

Perhaps it would be possible to widen the skip of test_strip_ycbcr_jpeg_2x2_sampling to take into account newer libjpeg_turbo? Currently it only skips 2.0, but we're on 2.1 (and ABI-wise, we're only pinning to 2, without the minor version).

Our GitHub Actions uses libjpeg turbo 2.1 and passes - https://github.com/python-pillow/Pillow/actions/runs/3819974382/jobs/6497949271#step:8:39

If you'd like that to be investigated further, please open a new issue.

[h-vetinari]

Yeah, I should not have responded so quickly (but wanted to avoid an inadvertent merge in that moment).

The 2.1 remains necessary on osx, where test_strip_ycbcr_jpeg_2x2_sampling does not fail on osx, and so I had actually read the right results.

Sorry for the back and forth 😑

[h-vetinari]

Our GitHub Actions uses libjpeg turbo 2.1 and passes - https://github.com/python-pillow/Pillow/actions/runs/3819974382/jobs/6497949271#step:8:39

If you'd like that to be investigated further, please open a new issue.

Fair enough. Since it's just a minor accuracy violation I'm going to skip that test for our builds.