Remove replication job supervisor
Use the scheduler as the job supervisor, since the scheduler is already a fancy
supervisor, with its own backoff logic, process monitoring, etc.
This simplifies the job starting/stopping logic and fixes a bug where the
simple_one_for_one supervisor could restart a job, but the scheduler would
consider it not running, and try to start another job with the same replication
ID on the same node. Since jobs register themselves in pg, the second job would
keep crashing with duplicate_job error the first time it tried to checkpoint.
8 files changed
