From Pixels to Structure: Lightweight Vision-Language Models for Document OCR and Structured JSON Extraction