Support parameter tags in optimizer weight decay

[dlwh]

Summary

• allow OptimizerConfig to tag parameters via TagPattern
• use parameter tags when computing weight decay mask
• support tag-based learning rate configuration through dict learning_rate
• implement scale_by_tagged_learning_rates to apply tag-based LRs
• document tagging features in the configuration guide with an expanded section on tag matching
• extend tests for tag priority and multiple tagged learning rates

Testing

• pre-commit run --files docs/reference/Configuration.md src/levanter/optim/config.py tests/test_weight_decay_mask.py
• pytest tests/test_weight_decay_mask.py -q

---

https://chatgpt.com/codex/tasks/task_e_685068f5119c8331ae1085b344c65a44