Move refreshCheckout out of db transaction

[finestructure]

This should address our analysis hangs. There is perhaps an issue upstream how db transactions are being handled in FluentKit and console-kit should perhaps not hang (reported here) but with this in place I believe we'll get rid of the service hangs that require manual intervention.

This has become more frequent recently, and I believe the reason is simply that we're seeing more new packages being added, which leads to a higher chance of a checkout (+ update work) taking us across the 10s event loop timeout.

Fixes #2227

[finestructure]

Another reason this has likely become more frequently noticed is the following: A few weeks ago I added an alert that explicitly fires when analysis doesn't log anything within 5 minutes (== hang detection). So now we get a "hang alert" pretty much immediately.

Previously, we only received alerts when there were no build reports after 4 hours (!) - with the underlying reason being a stalled analysis.

This long delay between issue and alert meant that we probably "solved" a number of those hangs by redeploying while the system was in a pending alert state, thus reducing the perceived incident frequency. (In particular on dev, where we deploy quite frequently.)