Make MosaicGPT a HuggingFace PreTrainedModel

[dakinggg]

(you can ignore the long list of commits this PR claims to have, it is just because I started working off of the 0.13.2 merge PR before it was merged to main)

This PR makes MosaicGPT a subclass of transformers.PreTrainedModel, allowing us to call generate and other HF utils. Note: there are still some todos, like supporting kv caching, but I think all of those should be called out with explicit errors in the code, and should be easy to add once this base is in. Note: this PR will wait for the KV caching PR to be merged to main before merging.

Testing todos:

• make sure generate works properly on a trained model, both with and without padding
  Shows 7b with batched (i.e. padded) and not batched output (i.e. not padded)

['I believe that MosaicML is the best company in the world. Someday, I hope to be a part of it.\nI am a senior at the University of Texas at', '<|endoftext|><|endoftext|>Q: What city is the capital of the United States?\nA: Washington, D.C. The capital of the United States is Washington, D.C.\n', '<|endoftext|><|endoftext|><|endoftext|>Q: What is the capital of the United States?\nA: Washington, D.C. The capital of the United States is Washington, D.C.\n'] 2023-03-20 22:53:23

I believe that MosaicML is the best company in the world. Someday, I hope to be a part of it.
I am a senior at the University of Michigan,

Q: What city is the capital of the United States?
A: Washington, D.C. The capital of the United States is Washington, D.C.

Q: What is the capital of the United States?
A: Washington, D.C. The capital of the United States is Washington, D.C.

• compare training run before and after, need to test with all three attention implementations
  (base run is from before this PR and uses flash attention)

• add tests for save_pretrained and from_pretrained
• add kv caching support after vitaliy's pr

[dakinggg]

1bs training again

[vchiley]

Chonker PR!

[vchiley]

TY for the updates!