feat: add metric evaluation and refinement for image2drawio pipeline

[ling0gyh]

Description:
Add post-processing evaluation and refinement stages to the image2drawio pipeline that detect
uncovered regions and rescue them as fallback image crops.

Changes:

• metric_evaluator.py: Content-coverage scoring with three-channel bad-region detection (fine,
  coarse, complex) plus banner channel for wide uncovered headers. Includes text-protection
  filtering, NMS deduplication, union-coverage check against existing elements, and debug
  visualization output.
• refinement_processor.py: Fallback rescue for uncovered regions — crops bad regions from the
  original image, filters out mostly-white regions, text-stroke regions, and regions already
  covered by existing elements, then returns them as image elements for the final XML.
• utils.py (sample_fill_stroke): Improved stroke color extraction — detects whether shapes have a
  genuine dark border vs anti-aliased edges, avoiding false black stroke colors on colored shapes
  (pink, teal, purple, etc.).
• utils.py (extract_text_color): Fixed inconsistent bbox clipping bounds.
• wf_paper2drawio_sam3.py: Integrated evaluate and refine nodes into the pipeline between
  build_elements and render_xml.