[FR] Use `server` to make inference faster

[NightMachinery]

whisper.cpp ships with a server. Isn't using that faster than loading the model again for each request?

Doing this should be much easier than https://github.com/natrys/whisper.el/issues/22.