fix: handle single-query exact search

[GioiaZheng]

Why

DenseRetrievalExactSearch.search() clamps torch.topk with len(cos_scores[1]). When the input has a single query, cos_scores has shape (1, n_docs), so reading row 1 raises an IndexError before retrieval can return results.

Closes #217.
Closes #170.

What changed

• Use cos_scores.shape[1] to clamp top_k by the corpus/document dimension.
• Add a regression test that runs exact dense search with one query and two documents.

Validation

• .venv\Scripts\python.exe -m pytest tests\retrieval\search\dense\test_exact_search.py
• .venv\Scripts\python.exe -m ruff check beir\retrieval\search\dense\exact_search.py tests\retrieval\search\dense\test_exact_search.py
• .venv\Scripts\python.exe -m ruff format --check beir\retrieval\search\dense\exact_search.py tests\retrieval\search\dense\test_exact_search.py

Risk / follow-up

Low risk. The change only corrects the top-k candidate limit to use the document dimension and leaves the ranking and heap logic unchanged.

[GioiaZheng]

Quick status follow-up: this remains ready for review and mergeable against current main. The change is limited to correcting the document-axis top-k clamp plus a single-query regression test. Happy to adjust the fixture style or rebase if preferred.