Make it possible to use env vars to downgrade models

Taylor's request: Use env vars to control models in production. This is because recently Fable was taken down suddenly and we had to re-deploy to change the model.

Upgrading models often requires changing the code more extensively, and different models behave differently and we should test and evaluate different model versions offline. We may need to use different models in different chat services and several different models for tools and other LLM-based tasks within services. Therefore, the env vars should be used to downgrade the main chat models when something goes wrong, and not as our primary way of setting the model for each LLM call.