Prevent negative ImageStat variance

[lllleolin-max]

On affected platforms, a constant white image of size (919, 405) can produce a slightly negative ImageStat.Stat.var due to floating-point rounding, causing stddev to raise a math domain error.

Current implementation (updated 2026-09-14):

Adopted @radarhere's suggestions from lllleolin-max#2, preserving all three original commits through e5203ab.

• Accumulate sum and sum2 as integers and use integer exponentiation, avoiding the rounding in this regression instead of clamping the result afterwards. Their return annotations are now list[int]; variance remains floating-point.
• Keep the regression focused on the affected constant image and assert var == [0.0].
• Simplify the release note accordingly.

Validation of this revision:

• 12 related ImageStat, ImageEnhance and histogram tests passed, with no skips.
• 781 temporary numerical checks passed, covering constant images, masks, RGB, histograms and nonzero-variance controls.
• selftest.py: 59 passed.
• Black, Ruff, mypy, Sphinx-lint and whitespace checks passed.

Local validation used Windows/Python 3.12.11 with this checkout's ImageStat and ImageEnhance modules and Pillow 12.3.0 prebuilt native extensions and remaining modules. Native extensions were not rebuilt; the full test suite was not run locally. This does not claim all floating-point variance rounding is eliminated. New upstream CI is pending completion.

Prepared with AI assistance; the adopted implementation and test simplification are @radarhere's original commits.

[radarhere]

I've found that the floating point error can be fixed by using integers. See lllleolin-max#2