feat(evaluations): integrate offline judges and deterministic scorers - #44

Summary

Adds the typed offline judge and deterministic scorer foundations for evaluations from code, stacked on #39.

• runs typed JudgeReference and Scorer values alongside each successful generation
• passes generation output plus row index, rendered input, expected output, variables, and metadata into evaluation logic
• resolves LaunchDarkly judges with pinned judge identity and diagnosable failure results
• avoids online metric emission from offline evaluations
• exposes the typed evaluations API from launchdarkly_ai_server and the convenience package
• documents and tests the judge/scorer integration without live LaunchDarkly or provider calls

Generation-result publishing and ingestion behavior is inherited from #39 and is intentionally not implemented or modified by this PR.

Stack

• Base: feat(evaluations): run generation-only evaluations from code #39 (devin/1786604824-evaluations-scaffold)
• This PR remains a draft while the evaluations stack is under development.

Validation

• uv run pytest packages/client/tests/test_evaluation_judges.py packages/client/tests/test_evaluation_scorers.py packages/client/tests/test_evaluations.py packages/client/tests/test_evaluations_run.py -v — 48 passed
• make lint
• make format-check
• make typecheck currently reaches the repository's pre-existing duplicate conftest module error.