Add initial_weights support for fine-tuning

[sirmarcel]

Summary

• Adds initial_weights setting to settings.yaml — loads model weights from a previous run's checkpoint (.msgpack) while starting optimizer, step counter, and data iterator fresh
• Recursive merge_params supports partial architecture matches (new layers keep random init)
• Records initial_weights path in saved config for reproducibility
• Fine-tuning example (my_experiment_finetune/) and tox smoke test
• Unit tests for merge_params

Test plan

• uvx tox -e tests — all 22 tests pass
• uvx ruff format . && uvx ruff check --fix . — clean
• Fine-tuning example runs end-to-end (DATASETS=.. lorem-train from my_experiment_finetune/)

🤖 Generated with Claude Code

[PicoCentauri]

LGTM!