[FEAT]: Add per-field confidence scoring and requires_review flag to LLM extraction

[SHANKAR8983]

Summary

In production emergency form processing, blind trust in LLM output is
dangerous. This PR adds heuristic confidence scoring to each extracted
field and a top-level requires_review flag so downstream systems and
human operators know when to intervene.

Changes

• src/llm.py: Added _compute_field_confidence() — scores each field 0.0–1.0
• src/llm.py: Added build_extraction_result() — wraps all fields with scores + _meta block
• src/llm.py: main_loop() now logs confidence-scored output
• src/test/test_confidence.py: 15 unit tests, all passing ✅

Example Output

{
  "name": {"value": "John Smith", "confidence": 0.9},
  "date": {"value": null, "confidence": 0.0},
  "_meta": {
    "requires_review": true,
    "low_confidence_fields": ["date"],
    "overall_confidence": 0.45
  }
}

Why this matters

• Flags incomplete extractions before they reach the PDF filler
• Integrates cleanly with existing structured logging (PR [FEAT]: Add structured logging across the project and replace print() diagnostics #365)
• Zero impact on existing add_response_to_json fix (PR Fix: prevent crash when appending to non-list JSON field #337)
• 15 unit tests covering all edge cases