Using pip is faster for macOS... running the tests takes ~45m to for x86_64 and ~15m for arm64 (!) compared to ~70 and ~30m, respectively, on main. I'm going to reenable slowtest tests for arm64 and see if it still comes in under an hour and keep it if so!

On main it's 28m for the pytest in the most recent green run for macos-latest with:

5116 passed, 42 skipped, 1191 deselected

On this PR it's a bit longer at 42m, but it tests more completely because I got rid of the not slowtest:

6193 passed, 47 skipped, 109 deselected

And the Windows one dropped from 60m to 34m for the pytest step, which is good, too. I realized it's now redundant with the Azure run, though, so I'm removing it (better to reduce GHA usage a little, and spread to Azure as another provider). The slowest test is now the most complete linux run (which hasn't really changed speed), which is fine!