Complete 'voice mode' implementation

[mattlindsey]

Complete tasks in #331 so we can turn this nice feature on!

Set VOICE_FEATURE to true in options.yml to try what's already been done.

[keithschacht]

Personally, this is the functionality I’m most excited about. I have built three different implementations of voice modes. HostedGPT was my to first but I was never happy with the latency. I prototyped another architecture and threw it out. And now I’m on a third implementation that I think might be the right track. If I get this new proof of concept fully working then I can figure out how to port it back since HostedGPT is a much more solid foundation.

I guess this is all to say: if you try out voice feature and like it, you can probably write a much shorter punch list than 331 in order to get out some kind of stable v1. And then I can help with a better rearchitecture when my current proof of concept settles down.