Improve egui_kittest screenshot change detector

[emilk]

Different hardware produce subtly different results:

https://docs.rs/egui_kittest/latest/egui_kittest/#what-do-do-when-ci--another-computer-produces-a-different-image

So we want a screenshot test that ignores small changes but detects large ones.

We should figure out a good default algorithm for this test by building up a test set of pairs of images that should either trigger or not trigger the threshold.

For instance, we recently discovered that these two images fell under our default threshold (i.e. were considered the same) for 98% of all pixels.

The screenshots above came up during the development of rerun-io/rerun#9151
That piece of work transitively broke the test in question and expected a failure but didn't get one.
It is important to note that while this was using egui_kittest's default per-pixel treshholds, it already set a "pixel-ignore-treshold" (as described in #5683) https://github.com/rerun-io/rerun/blob/6001edac74fb39d022eaeac310958a78ac791b4f/crates/viewer/re_view_spatial/tests/visible_time_range.rs#L417

Related to:

• Add pixel_threshold and improve threshold docs #5683

[lucasmerlin]

Maybe we can introduce a per-OS treshold?
Then we define one source OS where the treshold is really low. Ideally the snapshots for this OS would always be software rendered via some docker image (so the source OS would be linux), This way everyone, independent of their actual OS, can update the snapshots.

When testing for the other OSes the treshold could then be overriden in CI via some environment variable?
And in special cases where even more threshold is needed, the test could just set the treshold via some cfg(target = ) thing