Add built-in retrieval quality checks #2445

Overview

Add the standard retrieval-quality metrics (Recall@K, Precision@K, HitRate@K, MRR, NDCG@K, and InfAP) as first-class Check classes in the giskard.checks library. Currently, evaluation workflows (including the rag-evaluator skill) have to wrap each metric formula in a custom FnCheck, duplicating the implementation across user code and skill documentation. Promoting these to built-ins gives users a consistent, tested API for retrieval evaluation and eliminates a class of common implementation mistakes.

To-Do

• Implement six new Check classes in giskard.checks (e.g. giskard.checks.builtin.retrieval): RecallAtK, PrecisionAtK, HitRateAtK, MRR, NDCGAtK, InfAP.
• Each check accepts k where applicable, a threshold parameter for pass/fail, and JSONPath keys for the labelled relevant doc IDs and the retrieved doc IDs (e.g., relevant_ids_key, retrieved_ids_key).
• Support the three scoring strategies documented in the rag-evaluator skill's retrieval-metrics.md Section 4: Strict (default), Cosine similarity, and LLM-judged relevance. Strategy selection should be a constructor option, with sensible defaults.
• Add unit tests covering: empty relevant_ids, empty retrieved_ids, perfect retrieval, partial overlap, ranking sensitivity (for MRR / NDCG), and sparse-label cases (for InfAP).
• Add API reference documentation for each new check, including worked examples that mirror current FnCheck recipes for an easy upgrade path.
• Update the rag-evaluator skill (retrieval-metrics.md and examples.md) to use the new built-ins; deprecate or remove the custom FnCheck wrapper recipes once the built-ins ship.

Acceptance Criteria

• All six retrieval metrics importable from giskard.checks and usable directly in .check(...) calls without writing custom FnCheck wrappers.
• Each check supports configurable k, threshold, JSONPath keys for relevant/retrieved IDs, and a scoring-strategy option (Strict / Cosine / LLM-judged).
• Unit test coverage includes the edge cases listed in the To-Do above and is green in CI.
• Reference documentation is published on the Giskard Docs website, with worked examples for each metric.
• The rag-evaluator skill no longer ships inline FnCheck reimplementations of these metrics; it points users at the built-ins.