Refactor audio flows so that server streams audio to the client.

This way server has control over when to stop recording (in case clients
time out or don't start consuming bytes) AND it solves the attribution
problem (Voice Interaction Service would only get the capture session upon a positive
trigger, so would be "blamed" then).

Test: atest CtsVoiceInteractionTestCases
Bug: 168305377
Change-Id: If5bead87e88cfc9e31393029df5389afb6922183