Move the whisper call to a new module voip.ai, with new extra dependencies.

Next, we should add a full agent call class with Ollama to process the transcribed audio and Pocket TTS to create a voice response.

voip.audio should not need any of the extra ai dependencies.