Multi-model abstraction

[lalalune]

Right now we have a kind of ugly setup for switching between Llama, XAI, Claude, OpenAI and local.

We probably want completions to be a service, with more effort paid to moving it out of the runtime.

• make sure that the bot works with no API keys
• make sure it works with all OpenAI keys
• Add helper function which correctly figures out the endpoint, etc without needing to juggle in the .env
• Character file should configure which models are used where, and what company
• Maybe we have a "fast/cheap and slow/powerful" option for each provider so we can run like shouldRespond and other apis that get hit a lot for fast/cheap, and responses use slow/powerful
• generally clean up the response handling and make everything nice
• Make sure frequency penalty, presence penalty work for Grok, OpenAI, and repetition penalty for Llama
• add as LITTLE abstraction as possible. as few new classes or files as absolutely necessary. complexity is our enemy.