This increases CPU usage by ~2% for the audio_conv task, which I
think is an acceptable tradeoff. The Speex resampler still
makes its own allocations in internal RAM.

 - input files are upsamples and padded to stereo before playback
 - any in-progress playback is cancelled before playing a new file

This allows the audio pipeline to remain responsive even when the drain
buffer has completely filled. This in turn means that you now see the
track info in the 'now playing' screen change if the current track
changes whilst you are paused.

Since I was fucking around a lot in the audio processor anyway, I also
added mono->stereo expansion so that playing mono tracks on Bluetooth no
longer destroys your ears.

This is a more accurate way of knowing which track is playing when, and
also simplifies a lot of fragile logic in audio_fsm