Fix LLaMA model loading

Fix hot model reload with `drama_llama` backend not working #7

* Loading is now entirely in the worker thread.
* During load, the settings are locked but the brief blocking before is gone.
* Several new `Request` and `Response` messages were added.

The issue with the panic for unsupported models (on Metal at least) is still there. That requires changes in `llama.cpp` itself or duplication of code.