Antipattern will break legacy dataflow python pipelines if a new cy_combiner is added and used in the python counter_factory

There are two maps keyed by the cy_combiner object type in dataflow/internal/apiclient.py
structured_counter_translations and counter_translations 

The only way to make changes today without breaking the legacy python dataflow runner harness.  

(1) Introduce the new cy_combiner and modify the apiclient.py's maps which are keyed by the new types in a first PR
(2) Import and release the dataflow containers
(3) Add your code to use the new cy_combiner in the beam SDK

Prior to step 2, you cannot even run dataflow pipelines using your PR. As you will hit a KeyError when the cy_combiner type is looked up in the api_client maps.

Naturally a non dataflow contributor will just try to use the cy_combiner and unexpectedly break the legacy python dataflow runner harness.  

One solution is to catch the key error and log a warning that the metric is dropped, rather than failing the pipeline.

Imported from Jira [BEAM-7050](https://issues.apache.org/jira/browse/BEAM-7050). Original Jira may contain additional context.
Reported by: ajamato@google.com.