feat: Optimize LLM extraction from O(N) sequential calls to O(1) batch JSON request

[saurabh12nxf]

Description

This PR refactors the core LLM prompt generation in src/llm.py to dramatically improve performance.

Previously, the main_loop() function sent a brand new network request to the Ollama endpoint for every single PDF field. This was an O(N) operation that forced the local LLM to re-evaluate the entire incident transcript from scratch for each field, causing generation times of over a minute for larger forms.

I rewrote build_prompt() and main_loop() to accept the entire list of fields at once. By utilizing Ollama's format: "json" argument, we now force the LLM to extract all data natively and return it as a structured JSON object in a single O(1) batch request.

This drops generation time down to a fraction of what it used to be, speeding up the core application significantly.

Fixes #196

Type of change

• New feature / Performance Enhancement (non-breaking change which adds functionality)

How Has This Been Tested?

I tested this locally using llama3 as the backend model against a standard 7-field test PDF.

• Test A (Repro path): Generated a form with 7 fields. It successfully batched all 7 fields into a single JSON request. Output successfully mapped to the final answers_list array.
• Test B (Performance): Using a simple time.time() intercept, confirmed that generation dropped from ~70 seconds (7 sequential calls) down to ~17 seconds total (1 batch call).

Test Configuration:

• Firmware version: N/A
• Hardware: Windows Desktop / Local inference
• Python SDK: 3.12+

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes