Enhance learning rate scheduling flexibility

[benbennett]

Summary

• extend the scheduler factory to include warm restarts, exponential, and polynomial decay while exposing configurable thresholds
• clamp learning rates to optional minimums and improve logging when stepping multiple optimizers
• surface new command-line flags so both training entry points can tune the richer learning-rate policies

Testing

• python -m compileall .
• python -m compileall options

---

https://chatgpt.com/codex/tasks/task_e_68d3724865f08331b23931005646551f