fix: load classifier once per pipeline run

[SeanClay10]

Summary

The XGBoost classifier artifacts were being reloaded from disk on every PDF processed, which adds unnecessary overhead on batch runs.

• Moved load_classifier() from the per PDF worker function into run_pipeline() so the model, vectorizer, and encoder load exactly once per run
• Cleaned up dead code in _process_single_pdf() — removed leftover directory collection logic and an unused loop from a previous refactor
• Fixed flake8 linting errors in chunked_biomistral_llm.py and chunked_extraction.py

Closes #59