feat: unify image input into read: + an image: attach field (D.1 Phase 2)

[mirpo]

Unify the image glob into read: + a clean image: attach field

imagePath conflated three jobs in one field (self-glob + modulo row-index + attach). Phase 2 splits them into the same model everything else uses:

- name: images
  read: ./images/*.jpg          # enumerate files into rows
- name: describe
  forEach: images
  image: "{{.item.path}}"        # attach the current row's file as a vision image
  prompt: Describe this image. Return as JSON.

Changed

• New image: field on prompt steps — a templatable path rendered against the row (promptbuilder.RenderString, sharing buildValues with BuildPrompt), then base64-attached via the existing fs.ImageToBase64 → data:image/... path.
• Removed config.ImagePath + Step.HasImages, utils.setImagePath, the resolveIterations image-count branch, and fs.PickImageFile / fs.CountFiles / fs.globFiles (fs/images.go is now just ImageToBase64).
• The prompt+image are parsed together for field discovery, so {{.item.path}} loads from the read source.

Vision example migrated

Reads local images (ships 3 small sample .jpg) → forEach → image: — no HF download / magick, fully self-contained. Live-verified on qwen2.5vl:3b: 3 images → correct bar-chart / red-circle+blue-square / green-triangle descriptions with colors + tags, 0 retries.

Tests (TDD)

promptbuilder.RenderString; prompt-step image attach from a row path + missing-image failure; preprocess rejects image: on read steps; updated the affected existing tests. Full suite + golangci-lint clean.

Completes Block D.1 (file-in). Next: D.2 (non-JSONL output: csv/json/md).