BLD: assorted cleanups in release workflow

[jorisvandenbossche]

Some small follow-ups on #466:

• In shapely, we noticed that when using multiple docker images in a single job (through a matrix parametrization), only one of the images was cached (and you could see the same here), requiring to specify a scope to get proper caching.
• When testing, we don't actually need to checkout the source (that also avoids the cd ..)
• It seemed that the cache was not working for the MacOS x86_64 build. Not sure if that is related to having two jobs on the same macos-12 runner (although they use a unique cache key, so wouldn't expect that), but I saw that vcpkg now has experimental support for directly using the github actions cache (https://learn.microsoft.com/en-us/vcpkg/consume/binary-caching-github-actions-cache), so trying out this option. -> this seems to work only partially .. For the arm build, it only restores cached binaries that are not arm specific (maybe because the runner is actually intel mac, nor arm?)
• Removing some no longer necessary version restrictions in the requirements-wheel-test.txt file

[brendan-ward]

I think we want this to be our minimum supported version of MacOS, even if we're building on a newer MacOS runner. So this should probably be 13 since 12 is now past EOL.

[brendan-ward]

It seems like MACOSX_DEPLOYMENT_TARGET is getting ignored during the build because it looks like we are building wheels for only >= 14 (per the errors from when repairing the wheel). I don't recall this being an issue before, so it might be that we're not setting this when using the lukka/run-vcpkg@v11 action?

[brendan-ward]

Oops, I read that wrong. This is setting 14 as the minimum version we build for, but we were still expecting MACOSX_DEPLOYMENT_TARGET to be 12 in pyproject.toml.