Updates to hf model type modeling code

[haileyschoelkopf]

This PR modifies the Huggingface code to:

• Allows the seq2seq/causal backends to be selected automatically (--model_args backend="causal"). No change in functionality typically unless a user overrides the default backend (e.g. to get around MBart being listed as causal in transformers)
• Adds more typehints
• Modularizes HFLM.__init__()
• re-adds support for passing a transformers.PreTrainedModel through pretrained model arg
• warns the user this only works for single-process runs

Pretty big code change to HFLM, so maybe we should leave this staged for a day or 2 to make sure no bugs get introduced?

closes #1068 , #1084 .

[haileyschoelkopf]

Also addressing #1084 prior to merge.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ haileyschoelkopf
❌ zoryzhang
_{You have signed the CLA already but the status is still pending? Let us recheck it.}