FEAT add AgentThreatRulesScorer (ATR taxonomy scorer)

[eeee2345]

Adds AgentThreatRulesScorer, the scorer half of #1702 (the dataset loader landed in #1715).

What it does

• A deterministic TrueFalseScorer that evaluates text against the open Agent Threat Rules (ATR) ruleset via the pyatr engine.
• Returns True when at least one rule at or above a configurable min_severity matches; attaches matched rule ids, ATR category, and max severity as score metadata.
• Mirrors SubStringScorer's shape (true_false base, _score_piece_async, _build_identifier, min_severity validation).

Dependency

• pyatr (>=0.2.6, which bundles the ATR ruleset) is an optional dependency, imported lazily with a clear ImportError. The unit test uses pytest.importorskip("pyatr"); if you'd like CI to exercise it, pyatr can be added to the test extras — happy to wire that into whichever group you prefer.

Pairs with the _AgentThreatRulesDataset loader: the dataset supplies ATR-derived adversarial prompts, and this scorer detects whether a response trips an ATR rule.