[cookie-store] Deflake service worker registration

[Hexcles]

Instead of attempting to unregister previously installed service workers
at the beginning of a test, use add_completion_callback to unregister
them at test completion. This will deflake the tests when they are run
multiple times without restarts (--rerun).

The alternative is to do what many tests in service-workers do:
unregister service workers at the start, but wait for newly installed
service workers to become activated and use registration.active
instead of registration.installing, which is a bit more complex.

Fixes the flakiness observed in #17866 .

[Hexcles]

There are a few other tests in this directory that can benefit from the same treatment, but they seem to be flaky both with and without the patch (there's probably some other existing flakiness). I'll perhaps send out a part 2 PR for those so that this PR can be green.

[inexorabletash]

Seems fine.

I think we went with the "unregistered first" pattern to defend against leftover data from a failed/flaky test that didn't clean up, but if this improves things in practice then let's do it.

[Hexcles]

I think we went with the "unregistered first" pattern to defend against leftover data from a failed/flaky test that didn't clean up, but if this improves things in practice then let's do it.

That does work in principle. In practice, I've observed that registration of a service worker from scratch is initially in the "installing" state, whereas registration of a sw immediately after unregistering the same one is in the "active" state right away. Tests in service-workers always wait for the registration to become active, so either code path will work. We could also do the same here; and I'm not familiar with service workers enough to tell which approach is better.

In general, I think it's safe to assume that tests start from a clean state and strive to make sure tests clean up after themselves regardless of the results.