Some various fixes and changes for olmo

[hamishivi]

A bunch of changes. Mostly fixes to allow Olmo finetuning.

Additionally, I added:

• option to force bos token adding in training
• option to control distributed training timeout from the 30 minute default
• turned off safe saving due to it not working with our pinned accelerate version
• fixed the embedding check
• option for loading models with remote code
• eos token explicitly set in generation util
• ability to use HF_TOKEN env variable for model loading during evaluation
• sum loss instead of mean loss option for cases where you have to use lots of gradient accumulation.