Fix race condition in control loop that could cause pipeline starts to be ignored

[mwylde]

This PR fix a race condition in the controller that could cause configuration updates to be missed.

The potential race worked like this: the run_to_completion loop would read the config at the end of an iteration. If a JobConfig update arrived (via StateMachine::update) after this read but before the execute_state call in the next iteration, the JobContext for that next iteration would be created with the stale config. The StateMachine::update method would send a JobMessage::ConfigUpdate, but the state machine might have already proceeded with an older config.

This is addressed by adding a new Applied or NotApplied status to each configuration read from the database, which allows us to track whether we've started a controller loop with a particular configuration or not, preventing us from missing a configuration update.