Push metrics to a backend in an runner agnostic way

[kennknowles]

The idea is to avoid relying on the runners to provide access to the metrics (either at the end of the pipeline or while it runs) because they don't have all the same capabilities towards metrics (e.g. spark runner configures sinks like csv, graphite or in memory sinks using the spark engine conf). The target is to push the metrics in the common runner code so that no matter the chosen runner, a user can get his metrics out of beam.
Here is the link to the discussion thread on the dev ML: https://lists.apache.org/thread.html/01a80d62f2df6b84bfa41f05e15fda900178f882877c294fed8be91e@%3Cdev.beam.apache.org%3E
And the design doc:
https://s.apache.org/runner_independent_metrics_extraction

Imported from Jira BEAM-3310. Original Jira may contain additional context.
Reported by: echauchot.

[kennknowles]

The following subtask(s) are associated with this issue: #18595, #18596