fix: BEIR corpus _id and qrels mismatch when use_group_id_in_eval=True

[andrewwhitecdw]

What

Fix BEIR evaluation output when use_group_id_in_eval=True so qrels corpus-id values match the _id values written to corpus.jsonl.

The qrels writer now always uses the document ID mapping established while writing the corpus, keeping ID selection in one place.

Why

Previously, hash-ID mode wrote sequential corpus _id values such as d0, while qrels referenced hash IDs such as d_<hash>. Standard BEIR consumers therefore could not match relevant documents to corpus entries.

Behaviour

• Default mode continues to use sequential IDs.
• Hash-ID mode uses the content hash as both _id and group_id.
• Qrels always reference the corresponding corpus _id.

Validation

• Added a parametrized regression test covering both ID modes.
• GitHub lint, test, validate, and check jobs pass.

[oliverholworthy]

Thanks for identifying this. I added regression coverage for both evaluation ID modes and centralized qrels on the IDs written to corpus.jsonl. The hash-mode test fails against the previous implementation, while the default sequential mode remains unchanged. This restores the BEIR qrels-to-corpus identity contract.