Fix incompatibility with NumPy 1.20

[neutrinoceros]

Fixes an incompatibility encountered in the wild when importing Pillow==10.4.0 with numpy==1.20.x installed.

[radarhere]

NumPy < 1.24 is EOL - https://endoflife.date/numpy

Would an alternative solution be to upgrade NumPy?

[neutrinoceros]

NumPy < 1.24 is EOL - endoflife.date/numpy

One could argue that numpy < 2.0 is EOL since numpy devs have stated multiple times they cannot backport patches to 1.26 anymore, so I don't think NEP 29 (which I assume is the source for this endoflife page) is to be strictly followed.

The presence of this try/except block seems to indicate an intention to support versions older than 1.24, and it does work with any version but 1.20 , and it's very simple to fix that. I see that Pillow doesn't require numpy so I don't think it's possible to express an incompatibility in metadata (something equivalent to pip-style constraints files), except by requiring Python>=3.10 (numpy 1.21.4 being the first version to support it).
I can add this sort of information in a comment if it helps; my patch can be reverted when Python 3.9 is unsupported.

Would an alternative solution be to upgrade NumPy?

I should note that I discovered this in CI while explicitly testing a library with its minimal requirements (it has matplotlib pinned but not Pillow, since it's not a direct dependency). If this patch is undesired for some reason, bumping minimal requirements is my next best shot, but I'm emphasizing that it's not about a particular virtualenv that would be broken.

[radarhere]

One could argue that numpy < 2.0 is EOL since numpy devs have stated multiple times they cannot backport patches to 1.26 anymore, so I don't think NEP 29 (which I assume is the source for this endoflife page) is to be strictly followed.

That sounds like an argument for upgrading NumPy, not for staying on an old version.

my patch can be reverted when Python 3.9 is unsupported.

I'm mostly conscious of trying to have some line in the sand, so that we're not supporting old versions forever. If this request is just that we support NumPy 1.20 for the next year though, then that is a limited time frame.

[neutrinoceros]

If this request is just that we support NumPy 1.20 for the next year though, then that is a limited time frame.

Yes, I'm not planning to support Python 3.9 (hence, numpy 1.20) beyond early 2025.

Note that I don't think the new title for this PR is correct: this fixes a runtime crash, type checking isn't the only (or main) concern.

[radarhere]

this fixes a runtime crash, type checking isn't the only (or main) concern.

That is surprising - we install NumPy 1.17 in our Ubuntu 20.04 CI, and our test suite runs without a problem.

I'll go further then - neutrinoceros#1 will prevent numpy.typing from being imported unless the user is type checking.

[neutrinoceros]

That is surprising - we install NumPy 1.17 in our Ubuntu 20.04 CI, and our test suite runs without a problem.

I'm not surprised: numpy 1.19 and older or 1.21 and newer all work as intended: only version 1.20 has numpy.typing but doesn't have numpy.typing.NDArray.