TogetherSolver

[thesofakillers]

This PR contributes a TogetherSolver class, a solver for using models served by the Together AI API

Because Together supports the OpenAI python sdk, we simply create a subclass of the OpenAISolver, overriding some functionality. There is therefore some refactoring of the OpenAISolver included in this PR to facilitate this code sharing.

At the moment, we support the models specified in evals/registry/solvers/together.yaml, but in principle most models offered from the Together AI API can easily be added

Notes:

• logit biasing not supported by the Together API due to a lack of a unified tokenizer a la tiktoken from openai
• For the same reason, checking for context length limits not supported

[JunShern]

Thanks for adding this! This allows us to test lots more models including Llama, Mixtral, etc.