feat: Add LLM Confidence Validation and Human-In-The-Loop review (Fixes #222)

[RITVIKKAMASETTY]

What does this PR do?

This PR addresses the critical liability risk of the LLM silently hallucinating missing or ambiguous fields (like names, badge numbers, or incident codes) directly onto official PDF documents.

It implements a "Human-in-the-loop" validation pipeline by updating the LLM extraction to output structured JSON with confidence scores.

Changes Made

• Structured LLM Output: src/llm.py now uses prompt engineering to guarantee Mistral returns JSON ({"value": "...", "confidence": 0.95}).
• Confidence Thresholding: Fields with confidence < 0.85 are intercepted instead of blindly trusted.
• Fail-Safe PDF Generation: src/filler.py now maps values by explicit semantic field names, and writes [REVIEW REQUIRED] into the PDF for any low-confidence fields so responders can spot them instantly.
• API Handoff: Added a needs_review JSON column to api/db/models.py and the FastAPI response schema so the frontend can highlight flagged fields in the UI.

Testing Performed

• Added a full unit testing suite tests/test_llm_confidence.py covering high/low confidence branching, edge cases, and JSON parse failures (6 passing tests).
• Manually verified end-to-end extraction against ambiguous transcripts using the Ollama Mistral model.

Fixes #222