[airflow] Extractors should repeat input/output metadata in both START and COMPLETE

[rossturk]

There is an OpenLineage consumer that does not aggregate metadata across the lifecycle of a run, opting instead to ignore all but COMPLETE event types. This is reasonable for the use case of this consumer; they care about schemas and lineage but not operational status. However, it means that if lineage is provided during a START event, this consumer will miss it.

It would be ideal to accumulate metadata across the run on the consumer side, but this is currently not possible. For the time being, we can repeat input/output metadata and facet metadata in both START and COMPLETE events to accommodate.

[collado-mike]

This is the cause of some fairly frequent issues where an Airflow task fails before a START event is able to be sent (there seem to be a few paths to the on_task_instance_failed hook being called). The OL events can be quite large, so I understand the urge to reduce duplication, but unfortunately, it's causing some data to be missed entirely.

[denimalpaca]

This is already default BaseExtractor behavior, so we just have to go through the extractors and make sure that their extract_on_complete() do not just pass.