Baseer: A Vision-Language Model for Arabic Document-to-Markdown OCR