This removes speech recognition, pocketsphinx, training models
and the speech automated test interface.
This also re-establishes proper use of MOZ_WEBSPEECH to work
for the speech API (synthesis part only) that was a broken mess
before, with some synth parts being always built, some parts
being built only with it enabled and recognition parts being
dependent on it. I'm pretty sure it'd be totally busted if you'd
ever have tried building without MOZ_WEBPEECH before.
Tested that synthesis still works as-intended.
This resolves#1538
