Reports engine-ready, time to first word, end-of-stream flush latency,
real-time factor, and HWInference RSS, with the first session separated out as
it pays for the process starting up. Boot time is very small, cold vs. hot isn't
very large, but we want to know if we regress.
Audio is fed in real time over a MediaStreamTrack, so the real-time factor
comes from the engine's counters, not a wall clock the feed rate caps at 1.
Differential Revision: https://phabricator.services.mozilla.com/D324039