Engines

[daanelson]

WIP: Refactor of llama inference into separate engines which can serve predictions with different underlying inference code.

This has:

• proper config and predict.py setup (note - we can definitely really clean up our config here, pretty sure it can become a ~50 line .yaml file per model at most)
• code refactored into engine modules

This does not yet have:

• per-engine dependency management
• the aforementioned clean config
• more than one engine :D

[daanelson]

one thing that's coming out of this, at least imo, is that a lot of the config we're setting really lives at the engine level - that's especially true if we have special requirements for cog.yaml per engine, etc.

[technillogue]

are we still going to merge this?

[technillogue]

replaced by #47