Add retriever fine-tune recipe skill

[oliverholworthy]

Summary

• Add a combined retriever-finetune-recipe agent skill for Nemotron embedding and reranking recipes.
• Split family-specific guidance into embed/rerank/evaluation references and move recovery guidance into PITFALLS.md.
• Expose the same repo-local skill content through a .claude symlink without duplicating files.

Validation

• quick_validate.py .agents/skills/retriever-finetune-recipe
• uv run nemotron embed --help
• uv run nemotron rerank --help
• uv run nemotron embed finetune -c default -d
• uv run nemotron rerank finetune -c default -d
• uv run pytest tests/recipes/embed/test_cli.py tests/nemo_runspec/test_execution_uv_spec.py -q

Notes

• Distribution through Python wheels is intentionally left out until there is a proven skill-install/discovery strategy for pip-installed packages.
• Rerank torch-specific checks should be run in the GPU/Automodel environment before marking the PR ready.