[FEATURE] Support for minimum learning rate

[tmostak]

🚀 Feature

Support for specification of a minimum learning rate

Motivation

Often in the research literature minimum learning rates are set when fine-tuning a model using a cosine or linear schedule, preventing the learning rate from dropping to 0 by the end of training. It would be very helpful if this could be supported in the LLM Studio UX.

[pascal-pfeiffer]

There is actually an open issue for that at transformers: https://github.com/huggingface/transformers/issues/28441 Once merged there, we can quickly integrate in H2O LLM Studio, as we are already using the transformers schedulers.