new buffering and original features uploading logic [sound box testing needed]

[benejoseph]

Fix for two issues: features were being screwed up in tinytensor_features.c AND features uploading was not stopping the buffering of features immediately after keyword

• tested on my unit, should be tested in sound box too
• buffering issue was causing rippling / "poppy" features (see plots at end), fixed with memcpy circular buffer code + unit tests
• original features upload code back in. Tested live w/full suripu --> S3 loop and works.
• option to stream audio features with a #define
• for the unit tests, made the C++ protobufs be v2 instead of v3.

Features before and after (top is AFTER, bottom is BEFORE, note not the same sound going into the mic)

[benejoseph]

Proposal would be to go back to the net with which previous had a lot of false positives, and hope that this fix eliminates that issue.