DB connection in register phase breaks pre-5.4.1 behavior

[yaauie]

In v5.4.1, establishing the database connection pool was moved from the run-phase (where problems are retried) to the register-phase (which can cause a pipeline to fail to start).

This was done as a part of mitigating a connection leak caused by a new connection pool being instantiated for each invocation of the scheduler (an anti-pattern of the Sequel library we use).

The effect of moving from run-phase to register-phase is that if an instance cannot connect while a pipeline is starting, manual intervention is required to get the pipeline running again (e.g., restarting the Logstash process or modifying the pipeline definition so that auto-reload will kick in); the pipeline cannot self-recover from a transient error if that error occurs during pipeline startup.

---

I believe we should restore the prior behavior of establishing the connection pool in the run-phase, and to continue to avoid closing the connection pool until the plugin is closed (at shutdown). This would allow us to pull-forward the prior behavior of retrying initial-connection failures while avoiding re-creating the connection leak.