DataSource: Split off ffmpeg sniffer to a second pass, and only if necessary

ffmpeg's sniffer is intended as a grab-all-that-doesn't-work-elsewhere
extractor. Unfortunately, this causes two issues:

- As written, it requires ffmpeg to whitelist any extractors supported
by stagefright, or else it will blindly override them. This has codebase
sync issues, as shown by the VP9 and WAV accidental overrides
- It imposes an in-depth analysis of _all_ media, even that which we
want to be processed quickly by shipping stagefright plugins (hardware
or not). This is mostly noticeable in network streams and thumbnail
generation.

This patch moves FFmpegExtractor to an independent sniffer queue, and
it only invokes it when the regular MediaExtractor hits 1 of 3 conditions:

1 - The confidence in the identified container type is low
2 - No container is identified at all
3 - A video container was found by other extractors, but only 1 stream
(audio or video) was identified.

Change-Id: Ib96ff4f6bc06223fe0e819a57560d3c872a79ddd