Add section-aware chunked inference for resumes exceeding 512 tokens

Improve structured benchmark analysis and robustness

Clean up training pipeline and add export benchmarks