Add mode-aware structured evaluation

[ChathurangiShyalika]

This PR extends static_json evaluation for mode-selection scenarios.

Changes:

• Detect ground truth objects with one of response, clarification, or abstain.
• Score whether the model returns exactly one top-level mode key.
• Check whether the returned mode key matches ground truth.
• Extract required terms from ground truth answers and check term coverage.
• Add aggregate mode metrics to reports.
• Add tests for clarification, abstain, response, and noisy fenced JSON outputs.

Files changed:

• src/evaluation/scorers/static_json.py
• src/evaluation/report.py
• src/evaluation/tests/test_static_json_scorer.py