Document and unify engine audio & metadata retention

[drmfinlay]

Re: #194

CC @daanzu

The Kaldi engine has well-documented and flexible optional functionality for retaining audio and or recognition metadata for each spoken utterance. The natlink, WSR and sphinx engines also have functionality like this, but they aren't as well documented. There are also some differences between each implementation.

I've opened this issue for keeping track of unification and documentation of the retention features for each engine.

I can handle (at least) the sphinx engine changes.