While we're here, also add `brew tests --no-parallel` which I relied
on during testing.
Pretty much anywhere we rely on a stubbed formula on disk to work: we
need to disable the API.
These tests were very simple before and now this should result
in more code coverage without affecting test performance.
The only tricky thing was testing the `--missing` option without
actually installing a package using `install_test_formula` because
that is very slow (around 10 seconds on my machine). I ended
up just writing the tab to a plausible keg directory for each
package I wanted to "install". This allows us to test the behavior
while also not increasing CI time by ~20 seconds (though it'd
probably be faster on CI than my local machine).
