Region selection on label images to guide extraction

# Context

Extraction quality can vary from one uploaded image to another and with how each label is designed on the page (e.g. columns, tables, fine print). Users often know which parts of a label image matter most. Letting them select regions in the UI provides focused visual context the extraction pipeline can use as an improvement signal (e.g. crop, mask, or prompt augmentation), without replacing the full-image extraction path.

# Outcome

During the extraction flow, users can define one or more regions on each relevant label image. Those regions are included with the extraction request so backend/OCR/LLM steps can prioritize or scope processing. When no regions are selected, behavior matches the current full-image extraction path.

# Acceptance criteria

- User can select visible portions of label images in the UI as part of the extraction workflow (draw/adjust/clear selections).
- Selected regions are represented in a stable contract (page or image index, normalized or pixel coordinates) and sent with the extraction API request.
- Extraction remains correct and backward compatible when no regions are selected.
- UX makes it obvious what is selected and that selection is optional.