remove explicit install of Pillow throughout CI

[pmeier]

While looking into unifying our dependencies I wondered why we are installing PIL explicitly in multiple places throughout the CI, although installing torchvision should pull it in automatically. We aren't doing that with the other hard dependencies such as numpy or requests.

Turns out, we don't need to do that. After this PR, the only mentions of the Pillow version are

vision/setup.py

Lines 67 to 70 in b0dbbd7

	# Excluding 8.3.* because of https://github.com/pytorch/vision/issues/4934
	pillow_ver = " >= 5.3.0, !=8.3.*"
	pillow_req = "pillow-simd" if get_dist("pillow-simd") is not None else "pillow"
	requirements.append(pillow_req + pillow_ver)

and

vision/packaging/torchvision/meta.yaml

Line 31 in b0dbbd7

- pillow >=5.3.0, !=8.3.*

Given that these are "package blueprints" there is not much room of improvement. setup.py is executed, so we could parse the version from the conda recipe. But that feels a little to much IMO.

[pmeier]

This is a refinement as proposed in #5868 (comment). See the next comment for details.

[pmeier]

Initially, setuptools and similar packages used cp3_10 as the interpreter tag for Python 3.10. After PEP 641 was rejected, they changed to cp310. This patch was included in setuptools==59.1.0. With this tag set incorrectly, python setup.py develop ignores wheels on Python 3.10 and tries to use the source version. That is fine for pure python libraries, but means that it will try to compile Pillow from source.

Unfortunately, the defaults channel provides nothing between setuptools==58.0.4 and setuptools==61.2.0. The former has the problems described above and the latter has problems on Python 3.7 (see #5868). Thus, I opted to use the latest release in general and special case Python 3.7 here. When we hard-depend on Python 3.8, we can simply remove this.

[pmeier]

Failures are unrelated. mypy was fixed in #5970 and the ONNX break is going on for some days now.

[datumbox]

LGTM and the tests also pass (kicked them again to resolve the mypy problem).

@NicolasHug Any concerns? I think keeping the pil versions in 2 places is suboptimal but better than the previous approach, so I think it's worth merging this.

[github-actions]

Hey @pmeier!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py