storing 6 seconds (16Kbytes) of audio features for upload

[benejoseph]

We have a lot of false positives still on the okay sense, but when the desktop net is run on the uploaded 3 second feature buffer we get very few triggers. So if it's the case of the LSTMs getting into a "bad" state, a longer feature buffer ought to capture this better.

[benejoseph]

after upload: heap 38760 +: 50272 -: 10008

and then later heap 38704 +: 49592 -: 36512

[plasticchris]

so the low water mark is 10008 bytes, getting close to the limit