I've noticed a couple of flakes in TestPGCoordinatorDual_Mainline, e.g. https://github.com/coder/coder/actions/runs/5869096650/job/15913511403

I've identified a race condition during the startup of PGCoord that could be responsible.

In PGCoord's querier, we start pubsub subscriptions and workers on separate goroutines. That means that right at startup, we could:

1. get a mapping request
2. query for it and find no data
3. update hits the database, but we miss it because we aren't subscribed
4. subscribe to updates

This can only happen if something connects right as we are starting up, which is basically impossible in the field, but happens all the time in our tests.

The fix is to ensure the subscription happens before step 2, so that if an update comes in after 2, we will get notified by our subscription.

I can't be 100% sure this is the cause of the test failures, because we don't have logs that detail the ordering of these events. I've also added better logging so if something like this happens again we can debug.