Compare out of the box completions of Action Events on GGML MPT7B, LLAMA-2, GPT-3.5-Davinci, etc

[FFFiend]

Feature request

We basically want to see what model performs the best out of the box on generating action events given reference window state, ref actions and active window state, because we don't know what models we would like to add to openadapt.ml_models.provider other than GPT-3.5/4 for fine tuning

Refer to #327 for a bare bones, simpler example of testing completions purely on GPT-3.5/4, and stateful.py for how event dicts are currently sanitized before being passed into the prompt.

Also serves as a good exercise for using different ML API's (HuggingFace and OpenAI) and understanding how completions are made.

Motivation

Helpful for #379

[FFFiend]

Latest comment here has some LLAMA-2 models that run on CPU: #417

[FFFiend]

@KrishPatel13 I can get some work done on this if work with your other PRs is getting in the way, just let me know 👍

[KrishPatel13]

@FFFiend Sure no worries! I can work on something else 😊.

[FFFiend]

https://huggingface.co/conceptofmind/LLongMA-2-13b-16k

https://huggingface.co/conceptofmind/LLongMA-2-13b

LLONGMA models

[FFFiend]

https://github.com/facebookresearch/llama/issues/555 cant get LLama-2 to work inside a pipeline, not sure why.