fix(retrieval): preserve bi-encoder export metadata

[oliverholworthy]

What does this PR do?

Preserves bi-encoder wrapper settings in consolidated exports and restores them when the export is loaded again.

Changelog

• Persist retrieval task, pooling, and L2-normalization metadata on saved encoder configs.
• Restore metadata from config.json or the AutoModel v5 sidecar while preserving explicit loader overrides.
• Register bidirectional Llama and Ministral configs with Hugging Face auto classes.
• Add a helper that re-exports older HF-loadable bi-encoders with explicit retrieval metadata.
• Add focused registry, retrieval-loader, and bi-encoder tests.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added focused CPU-compatible tests.
• Added the migration helper required by the documented mining workflow.
• All commits include DCO sign-off.

Validation

• uv run ruff format <changed files>
• uv run ruff check --fix <changed files>
• uv run pytest tests/unit_tests/_transformers/test_registry.py tests/unit_tests/_transformers/test_retrieval.py tests/unit_tests/models/bi_encoder/test_llama_bidirectional_model.py -q (63 passed)
• uv run python examples/retrieval/data_utils/export_biencoder_with_metadata.py --help

Additional Information

Extracted from #2306 so that the fine-tuning documentation PR does not carry model-loading behavior.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.