Feat/xgboost rework

[raymondcen]

The biggest change is that PDFs can now be processed in parallel, which cuts down wait time significantly when running batches. There's a solid round of improvements to how text gets cleaned and how the AI extracts information.

Processing multiple PDFs at once

• Previously the pipeline handled one PDF at a time. Now you can pass --workers N to run N files simultaneously. It defaults to 1 so nothing breaks if you don't touch it.

Cleaner text going into the model
Spent time improving what actually gets fed to the AI. Noisy stuff like DOIs, figure captions, and numbered references gets stripped out. Weak or irrelevant paragraphs get dropped entirely using a scoring system, and section headers are now preserved correctly (they were getting removed before).

Switched to qwen2.5:7b, added retry logic for null responses, and bumped up the context limit so longer papers don't get cut off.

Misc fixes and optimizations

• Spell checker loads once at startup instead of on every call
• XGBoost automatically uses GPU if one is available
• OCR bypass option during parallel runs since it was causing workers to freeze

[SeanClay10]

LGTM!