Follow on from #8972.

This runs tests by default with pytest-xdist, except disables for valgrind.

Locally, tox -e py314 goes from 44.79s for pytest to 22.78s.

And we save about an hour on the CI:

workflow	before	after	saving	faster
test	46.7m	28.9m	17.8m	1.62x
test-docker	133.7m	102.0m	31.8m	1.31x
test-windows	19.1m	13.1m	6.0m	1.46x
test-mingw	1.7m	1.0m	0.7m	1.73x
Total	201.2m	145.0m	56.2m	1.39x

This means the test order is no longer deterministic, so we don't need pytest-reverse.

Another option is to add another "extra" or "dependency group" to install pytest-xdist/pytest-sugar for certain invocations (and it'll need some config only to only set the options for these).

I put the imagegrab tests in a single group, which means they'll be allocated the same runner, to avoid some resource clashes on the CI. And I put an AVIF test into its own group because it sometimes failed locally when my machine didn't have much memory to spare.