Detect text bounding boxes inside a polygon on a page

Crops the page to the polygon's bounding rect (with padding for PP-OCRv5 context), runs text detection on the crop, and filters detections by polygon centroid. One GPU round trip. Returns detected bboxes in page-pixel coords. Recognition is not performed; callers needing text content route the bboxes through /ocr-regions. Results are ephemeral and not persisted.