Rough plan
1. detect when this occurs (currently not distinguishing NNR vs NR).
2. record to media capabilities database
3. incorporate into smooth=true/false predictions alongside dropped frames


We want to detect NNR's irrespective of which stream (audio vs video) has underflowed. Both should be treated as a symptom of poor video decode performance (audio decode generally thought of as free, but can be starved by expensive video operations on the media/ thread).