fix(retrieval): support Sentence Transformers v6 metadata

[oliverholworthy]

Summary

• accept the Sentence Transformers v6 Normalize module path when it represents final sentence-embedding normalization
• reject v6 Normalize configurations that target token embeddings or redirect output, which the NeMo single-vector wrapper cannot represent
• keep emitting the v5.4-era module path so exported checkpoints remain loadable by Sentence Transformers 5.4 through 6.0
• update the lockfile to Sentence Transformers 6.0.0 so the default test environment exercises the new release

Follow-up to #3546.

Compatibility rationale

Sentence Transformers v6 moved Normalize to sentence_transformers.base.modules. The v6 release explicitly keeps existing models loadable, while checkpoints saved with the new path do not load on older releases. Keeping the older path for exports and accepting both paths on import therefore provides the widest compatibility.

Release notes: https://github.com/huggingface/sentence-transformers/releases/tag/v6.0.0
Migration guide: https://www.sbert.net/docs/migration_guide.html

Testing

• ruff format --check and ruff check on changed Python files
• 39 focused metadata and locked-v6 round-trip tests
• all 8 pooling and normalization round-trip variants on sentence-transformers 5.4.0
• all 8 pooling and normalization round-trip variants on sentence-transformers 5.6.0
• all 8 pooling and normalization round-trip variants on sentence-transformers 6.0.0

[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.

[akoumpa]

/ok to test a975dae