Strip Down OpenLM for a Clean Start

[drothermel]

Wanted to go through and remove things we are unlikely to use that make the code more complex.

The removed things:

• Everything github workflow and pre-commit related
• plots/ (used to demonstrate instruction tune perf)
• sagemaker_train/ (we're not using sagemaker)
• Everything MOE: [PR 115: Add MOE Support](Mixture of Experts mlfoundations/open_lm#115)
• Everything Averaging: [PR 238: Adding Averaging of Iterates](https://github.com/mlfoundations/open_lm/pull/238/files)
• Everything Mamba
• experimental_meta_device

Then:

1. Tried to get a dev environment using rye and then separately trying with conda, leading me to remove xformers which I just couldn't get to work.
2. Concluded that llm-foundry is crazy crazy bloated and composer is kinda confusing so my vote is we write our own eval loop (the only thing composer is being used for). Removed llm-foundry and mosaic deps from open_lm for now.
3. After removing all of that, I got a dev environment with rye that installs & then fixed formatting and lint issues!

Next Step: try to get something to run!