Add metrics dashboard: scorer quality and benchmark leaderboard

[hannahwestra25]

What this adds

Metrics dashboard (doc/dashboard/) — a new top-level docs section:

• 0_dashboard.md: landing page explaining what's here, what's planned, and how to refresh the data.
• 1_scorer_quality.py/.ipynb: an Objective Scorer Leaderboard (accuracy/F1/precision/recall) and a Harm Scorer Leaderboard (MAE, Krippendorff's alpha), rendered from the existing committed registries under pyrit/datasets/scorer_evals/. No new dependencies — reads local JSONL only.
• 2_benchmark_leaderboard.py/.ipynb: a Technique / Adversarial Model Leaderboard (success rate, N, success/failure/error/undetermined counts), rendered from the new committed store below.
• Both pages registered in doc/myst.yml's top-level TOC, after scanner.

Benchmark exporter identity fields + store (build_scripts/export_adversarial_benchmark_result.py):

• Every technique-metrics row now carries objective_target, objective_scorer, and dataset identity fields (added to JSON/CSV/TXT output).
• New --update-benchmark-store [--benchmark-store-path] flag that upserts rows into a new committed store, pyrit/datasets/benchmark_results/adversarial_benchmark_metrics.jsonl, keyed on (technique, adversarial_model, objective_target, objective_scorer, dataset). This is what the new Benchmark Leaderboard page reads.

Demo data behind the Benchmark Leaderboard: three rows from one real, live AdversarialBenchmark scenario run — role_play_video_game, context_compliance, and red_teaming techniques, one adversarial target, harmbench dataset, --max-dataset-size 1. This is genuine exporter/scenario output, not synthetic data (see caveat below).

Explicitly out of scope for this PR

• No ADO pipeline/YAML changes.
• No native objective-target sweep (a leaderboard comparing target models against a fixed adversarial model) — the exporter's identity fields support it, but building it is future work.
• No shared PR-opening credential automation.
• No CI automation that runs the benchmark scenario or refreshes the store automatically — both dashboard pages are refreshed manually today (documented in 0_dashboard.md). The existing manually-triggered .azuredevops/adversarial-benchmark.yml pipeline could in principle drive this, but today its export step never passes --update-benchmark-store, and it has no step to regenerate notebook outputs or to commit/push/open a PR (it has no GitHub write credentials). Wiring that up is a good follow-up, best done as its own change rather than bundled here — it also touches the same YAML file Improve adversarial benchmark dataset and scoring #2551 is actively modifying.

Known limitations

• Demo-scale data, not a rigorous benchmark. The Benchmark Leaderboard's data comes from a single small demo run (1 dataset item per technique, so N=1 per row) using one adversarial-model identity that happens to share a deployment with the objective target (no independently-configured ADVERSARIAL_CHAT_* target was available in this environment). Treat the page as a working demonstration of the mechanism, not a statistically meaningful or cross-model comparison. Broadening this into a larger, regularly-refreshed sweep across real independent adversarial models is future work (see the page's "Note on scope" section).
• The exporter logic (_objective_identity, _dataset_identity, _build_technique_metrics, _upsert_benchmark_metrics) is covered by unit tests using hand-constructed ScenarioResult/AttackResult/TargetIdentifier/ScorerIdentifier objects, and has now also been exercised end-to-end against the real scenario result described above — so both the isolated logic and the full memory-to-JSONL path are validated against genuine PyRIT output.

Testing

• pytest tests/unit/build_scripts/test_export_adversarial_benchmark_result.py — 12 new tests, all passing.
• pytest tests/unit/build_scripts tests/unit/common tests/unit/datasets — no regressions (268 + 5069 passed, 2 pre-existing skips).
• ruff check / ruff format --check — clean.
• python -m build_scripts.validate_docs — 163 TOC references validated, no orphaned files.
• pre-commit run against all changed files — clean, aside from two confirmed pre-existing, environment-specific issues unrelated to this change: check-added-large-files fails against any file (including already-committed ones) due to a git check-attr subprocess error in this sandbox; nbstripout reports "modified" on every run even though a before/after hash comparison shows the file is byte-identical.
• Ran a real, live AdversarialBenchmark scenario end-to-end (pyrit_scan run → export_adversarial_benchmark_result.py --update-benchmark-store → the new dashboard page) to validate the full pipeline against genuine data rather than mocks.