fix(rerank): accept sparse indexed results

[huangruiteng]

Motivation

OpenAI-compatible rerank providers may return sparse or top-N results with valid input indexes. The client rejected any result-count mismatch before reading those indexes, so the hierarchical retriever discarded valid rerank output and fell back to vector scores.

Fixes #3090.

Implementation

• Accept sparse result lists from the OpenAI-compatible rerank endpoint.
• Map each returned score back to its original input index while leaving missing documents at the existing zero default.
• Preserve validation for missing or out-of-range indexes.
• Downgrade sparse-result length reporting from warning to debug.
• Convert the previous mismatch-rejection test into an out-of-order sparse regression test.

Validation

• 22 focused OpenAI rerank tests passed; 2 unrelated historical RerankConfig assertions were deselected after both reproduced on the pinned base revision.
• 14 adjacent rerank unit tests passed.
• Ruff check passed.
• Ruff format check passed.
• Python compile check passed.
• Diff hygiene passed.

The two pre-existing base failures are:

• TestRerankConfig.test_default_provider_is_vikingdb
• TestRerankConfig.test_unknown_provider_raises_value_error

Risk and gaps

The behavior change is limited to the OpenAI-compatible rerank client. Full-length responses and other rerank providers are unchanged. Missing sparse entries receive 0.0, so existing threshold and sorting logic can naturally exclude them. Validation uses a mocked provider response matching the reported shape; no live provider credentials were used.

[huangruiteng]

@MaojiaSheng could you please review this focused fix for #3090? It corrects sparse OpenAI rerank result mapping and adds regression coverage; all current CI checks are passing. Thank you!