Implement Multi-Process Handling (#7)

[mostlygeek]

This change adds functionality for llama-swap to manage multiple llama-server backends. This is an advanced use case that requires enough GPU resources to load multiple models. An example is when programming and a larger model like qwen-coder-32B is used for code generation and qwen-coder-1.5B is used for auto-completion.

This implements the design from issue #7. A new groups configuration is introduced:

models:
  "qwen-coder-1.5b":
    cmd: llama-server --port 9001 -m models/qwen-coder-1.5b.gguf
    proxy: http://127.0.0.1:9001

  "qwen-coder-32b":
    cmd: llama-server --port 9002 -m models/qwen-coder-32B.gguf
    proxy: http://127.0.0.1:9002

groups:
  coding:
    - "qwen-coder-1.5b"
    - "qwen-coder-32b"

When a request to /v1/chat/completions with a {"model":"coding/qwen-coder-1.5b"} is sent llama-swap will load both models into the GPU(s). When a request for "coding/qwen-coder-32b" is requested the request will be served immediately. This can greatly reduces the time to first token for a request for both models in the configuration.

It is up to the operator to ensure the group configuration is viable. Generally watch out for these things:

1. there is sufficient GPU resources to load all models
2. there are no port conflicts in the configurations

[mostlygeek]

Renamed groups to profiles as this makes more sense when talking about it.