fix(eval): handle variable retrieval depths in MKQA recall

[BingH225]

Summary

• Size MKQA recall buckets using the maximum retrieval depth across queries.
• Add a regression test for queries with different numbers of retrieved contexts.

Problem

evaluate_qa_recall currently initializes its recall buckets using only the
first query's context count. If a later query has more retrieved contexts and
its first answer occurs beyond that initial depth, the hit is silently omitted.

For example, with retrieval depths [1, 2] and the second query matching at
rank 2, the previous implementation returns:

[0.0, 0.0]

The expected Recall@1 and Recall@2 values are:

[0.0, 0.5]

Testing

• Added test_evaluate_qa_recall_supports_variable_retrieval_depths.
• Executed the regression scenario directly against the modified module.
• Python AST parsing passed for the implementation and test.
• git diff --check and targeted codespell passed.

The complete pytest invocation was not run locally because the available
Python environment does not include FlagEmbedding's torch dependency.