Fixes flake seen here:

https://github.com/coder/coder/actions/runs/10677025332/job/29591518118

The original test waits for DB calls to update success and failure of notifications, and blocks on apause channel. However, it uses a single pause channel for both DB calls, and so implicitly assumes that both success and failure updates occur during an update sync.

However, the update sync timer isn't synchronized to anything, and so there is a race where the update sync only has either the success or the failure result but not both. This blocks the test, which waits for both, and deadlocks before unpausing.

It appears the unpause mechanism is there to test that thePendingUpdates metric updates accordingly.

This fixes the flake by:

1. using a Quartz clock to control when theManager syncs updates.
2. waiting for thePendingUpdates metric to reach 2, so that we know that both success and failure have been queued up
3. triggering the update via the Quartz mock clock

This has the nice property that a "pause" function is no longer required: we know the manager is in our desired state before we trigger the update, and can assert on the Metrics before and after.