fix memory leak

[jeremyandrews]

Currently each time we convert audio to text the kakaia process grows by a few hundred kilobytes. Need to track down where the memory is leaking, and fix.

[jeremyandrews]

The leak is because we're loading the model into memory each time we convert an audio file to text, where we should only be loading it one time. The logic in src/speech.rs needs to be reworked to only load the models into memory one time and then re-use.

[jeremyandrews]

Fixed in e5247c9

We now load the deepspeech::Model only one time at launch and share it to all threads with a mutex.