add support for tracking speech markers in bidirectional streams

[davehorton]

We should have some feature like twilio's mark event for bidrectional streams, to allow the remote party to track roughly which pieces of speech have been played. The idea is that while sending bidirectional streaming audio, they can intermittently -- perhaps at the end of each sentence for example, send a mark message and then we send a play_done message back when the audio to that point has been played out.