All line detection and splitting is now done in a separate LineBuilder
class. Lines which require OCR are later picked up by OCRBuilder

Character level splitting still not working perfectly

Line counts changed because of different behavior of new line det model
Picks up some lines within figures which weren't before, hence the
increased count. Outputs are still unchanged

Enables skipping of texify predictions if required

Skip on a few conditions, retain original text

Add another heuristic to clean up line numbers

Add Inline Math Support

Factor out llm services, enable local models