Stash WatchedItem changes so that the jobs run from the queue
Previously, the jobs ran at post-send instead of via runners.
The stash is still written to immediately and will quickly
replicate to all datacenters. Since the stash and DB data is
merged on access, users should still almost always see their
changes on the next request. There is still the small chance
of unusually high latency at the in-memory stash layer, which
should be rare.
Also inject the JobQueueGroup into the store instance instead
of adding more singleton references.
Bug: T188801
Change-Id: Ic1c2b5a3592469b5b8386012a9c3365fdcf5b8e1