Add manylinux_2_28 ppc64le wheels

[radarhere]

#9107 requests ppc64le wheels.

While there is the potential for using self-hosted runners to generate these quickly, if we don't want to go down that route, then here is another option - using QEMU to create them, but with minimal dependencies to reduce the build time. In the build in this PR, ppc64le jobs are not the longest.

[sandeepgupta12]

Just following up on regarding the ppc64le wheels. Are there any updates or feedback needed to move it forward?

[radarhere]

Simply waiting for review from another Pillow team member.

[aclark4life]

@sandeepgupta12 Can you test these wheels before we merge ?

[sandeepgupta12]

Sure, I can test the wheels on a ppc64le machine and let you know the results.

[sandeepgupta12]

@radarhere
I tested these wheels on a native ppc64le machine. The build works, but I see some test failures:

FAILED Tests/test_file_iptc.py::test_open - AssertionError: assert [_Tile(codec_..., args='raw')] == [('iptc', (0,...'raw', None))]
FAILED Tests/test_file_iptc.py::test_layers[3-RGB] - AssertionError: assert 'R' == 'RGB'
FAILED Tests/test_file_iptc.py::test_layers[4-CMYK] - AssertionError: assert 'C' == 'CMYK'
FAILED Tests/test_image.py::TestImage::test_show - Failed: DID NOT WARN. No warnings of type (<class 'DeprecationWarning'>,) were emitted.

Looking at the build script, on ppc64le the build function runs build_libjpeg_turbo but then exits early:

if [[ "$AUDITWHEEL_ARCH" == "ppc64le" ]]; then
    return
fi

Because of this, other dependencies are not built, so the resulting wheel has incomplete image format support. That seems to explain the above test failures.

[radarhere]

I would be surprised if the lack of image support was the reason for the test failures. The test failures sound more like #9088, which was merged after the last time I updated this PR. So I suspect the problem is that you built from this PR, and then tested main.

To simplify things, I've now merged main into this PR again, so that change should now be included. Could you try building and testing again?

[sandeepgupta12]

@radarhere @aclark4life Thanks for updating the branch. I retested on a native ppc64le machine, and everything is working fine now ✅. I think we can go ahead and merge this PR.

[hugovk]

How long does this take to build? How many wheels does it create? How big are they?

[radarhere]

Oops, accidentally deleted this branch. I've recreated it in #9216

The longest run is 1h 1m. This creates 7 manylinux_2_28 wheels (all active CPython versions, including threaded). They are each 1.9mb.