fix(evaluators): Update Luna support

[mikebranc]

Summary

https://app.shortcut.com/galileo/story/63852/agent-control-luna-guardrails-not-working-as-expected

• Fix metric names in Luna2Metric: output_toxicity → toxicity, pii_detection → pii; add input_pii and not_empty operator
• Require stage_name for local stages: previously only validated for central stages, but the API needs it for both — validator and field description updated accordingly
• Add categorical operator support: Protect's local-stage rule engine doesn't evaluate not_empty/any server-side (always returns not_triggered); added _evaluate_metric_results for client-side fallback
• Pass stage_name in local stage API calls: was being configured but not forwarded to invoke_protect
• Test fixes: updated all local stage configs to include stage_name, corrected stale metric names (output_toxicity, pii_detection), added test_local_stage_requires_stage_name and TestLuna2BuildMessage coverage

[codecov]

Codecov Report

❌ Patch coverage is 41.37931% with 34 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...agent_control_evaluator_galileo/luna2/evaluator.py	37.03%	34 Missing ⚠️

📢 Thoughts on this report? Let us know!