Improve RMSE math and lower thresholds in Canvas tests

[HalfWhitt]

Fixes #4005; I've been nerd-sniped. Uploading this with threshold values that work on macOS; presumably at least a couple will need bumping up a little depending on test failures.

Also removes the unused assert_pixel.

PR Checklist:

• All new features have been tested
• All new features have been documented
• I have read the CONTRIBUTING.md file
• I will abide by the code of conduct

[HalfWhitt]

First of all, I know most of the core team's on holiday leave, so I'm aware this might not get reviewed for a while.

@corranwebster, I'd appreciate a double-check on my math. I've changed how it's written out Python-wise, but math-wise the only differences should be, per your suggestions:

• Dividing by the four bands
• Converting the images to RGBa to test premultiplied alpha

Evidently the premultiplying helps a lot in some cases. After all, dividing by bands should, alone, only halve the error amounts... but the threshold for the test_multiline_text was able to drop all the way from 0.09 to 0.01, and test_transparency from 0.1 to 0.01.

test_write_text (which uses a variety of fonts, not just system as in test_multiline_text) only dropped from 0.07 to the expected 0.035, but it's interesting that Gtk on Wayland was the only one to go above 0.02.

[corranwebster]

Looks good, as a non-maintainer.

[HalfWhitt]

Looks good, as a non-maintainer.

Maintainer or not, I very much appreciate your canvas-and-image-related expertise so far!

[johnzhou721]

Nice use of ImageChops.difference!

[HalfWhitt]

Nice use of ImageChops.difference!

Credit to corranwebster, they're the one who suggested it 👍

[kattni]

@HalfWhitt Please ping me when this is ready for review. Thanks!

[HalfWhitt]

@kattni Ping : )

[HalfWhitt]

(This is still ready to go, I just decided to arrange it slightly differently.)

[freakboy3742]

Can't argue with much here - ImageChops.difference() is a new one for me, but I'm very much in favor of adopting something that is a pre-existing difference measure than inventing one. And, if it means our error thresholds are lower, all the better.

[HalfWhitt]

ImageChops.difference() is a new one for me, but I'm very much in favor of adopting something that is a pre-existing difference measure than inventing one. And, if it means our error thresholds are lower, all the better.

Just for the record, ImageChops.difference is only a tidier way of doing the same math as what I initially submitted:

        total = sum(
            ((actual - expected) / 255) ** 2
            for actual, expected in zip(
                chain(*scaled_image.convert("RGBa").getdata()),
                chain(*reference_image.convert("RGBa").getdata()),
                strict=True,
            )
        )

It's not relevant to lowering the error; that comes entirely from (1) dividing by the number of bands, and (2) premultiplying the alpha, so we're not testing invisible differences. (Both of which were corranwebster's idea on the original ticket).