Support generic models and arguments

[jlamypoirier]

• Use transformers factory methods to support all (causal LM) model types instead of a pre-defined list.
• Support any configuration parameter instead of a predefined list and explicitly validate them. Arguments are of the form key=value, with value type inferred automatically (not fool-proof but it should be good enough).
• Support pre-trained models.
• Add git-lfs to the image to help downloading big models
• Replace remaining check_unused usages with proper checks.
• Improve initialization/load time benchmark, measure individual components.
• Use inference mode (small speedup)
• Pass arguments explicitly to methods instead of the namespace.
• Fix generated token count.
• Remove unnecessary args for deepspeed.
• Update transformers.
• Update and trim makefile commands. Also serves as examples.
• Update shell scripts, mark python scripts as outdated.