[Feature request] Support learning rate warm-up steps and schedulers

Fine-tuning LM and co-training involving LM often requires LR schedulers. It would be nice to support various LR scheduler to reproduce SOTA results.

The issue names no files, tests, or entry points. Start by surveying the existing LM fine-tuning and co-training paths, then clarify which warm-up behavior and schedulers are required. Done should mean the agreed scheduler set and warm-up steps are supported and covered by tests, but the request does not define either scope.

Written by the indexing model from the issue text.