Improve model parametrization

[rvashurin]

Several improvements here:

• Model config makes use of composition feature of Hydra reducing duplication between example experimental configs.
• Support for device_map=auto feature of accelerate by removing all .to(device) calls to the model.
• Remove some obsolete model parameters like use_auth_token.
• Remove hard-coded limits on total sequence length

Completes following Notion tasks:

• https://www.notion.so/max-length-b5692b413f554b76b7c77863c69e5d0b?pvs=4
• https://www.notion.so/device_map-auto-1f1a6c9411b54bac83328793cc8b2ad7?pvs=4

[ArtemVazh]

I suggest removing the max_length parameter everywhere since we override it with max_new_tokens. Also, I suggest removing min_length in greedy_tokens.py or changing it to min_new_tokens, since it also overrides with a value equal to the input length.

[rvashurin]

@ArtemVazh max_length removed, but I suggest we do min_length stuff separately to not overload this PR with concepts and code.