test_empty_image: measure CPU time between filling and not filling

[akx]

Alternative to #9957.
Refs akx#21 (comment).
Refs #9933 (which fixes the timeout flake by reducing the size of the image, so the allocation is smaller, which was not a preferred approach).

[akx]

cc @radarhere :)

[radarhere]

While I instinctively dislike reducing the severity of a test case,
and my first instinct was to try and see if Pillow could overcome it with improved performance,
I do think this test is about measuring the overall performance for the scenario, not just whether we have mitigated the problem at all.

It's just unfortunate that there's such variety when testing in CI.

I've approved #9933.

[akx]

This is the more accurate test for 91b01f4 though: we're now measuring whether ImagingFill (which gets called for a non-None initial fill) actually bails out early when either dimension is zero. Without instrumentation at the C level, I don't think we can do much better than to measure it like this... 🤔

With this, the severity of the test case (100M x 0, or 0 x 100M) can remain as it was, with the noted caveat that the tall test will temporarily allocate 700M of memory for the row pointers.