feat(data): pluggable label-format adapters (COCO/YOLO/VOC ingestion)

[stanlrt]

Summary

Adds pluggable label-format adapters so users can point data.labels.format
at standard annotation files and RAITAP converts them to its native label shape
before alignment. No more hand-reformatting COCO/YOLO/VOC labels. Closes #338.

• New LabelFormat enum (native/coco/yolo/voc) + LabelsConfig.format
  field. native is the default and a no-op (zero behaviour change).
• Lightweight adapter registry + @label_format decorator +
  resolve_label_format_adapter, mirroring task_families/registry.py.
• Adapters: COCO (detection + classification), YOLO (detection),
  Pascal-VOC (detection). COCO/YOLO category ids pass through unchanged
  (preserve the model's label space); VOC class names map via
  model.class_names order, else the standard 20-class VOC order.
• Dispatch wired into DetectionFamily.load_labels and
  load_classification_labels; both reuse the existing id-alignment logic via
  an extracted _align_detection_records helper.
• Built via the spec-driven subagent workflow: per-task implement + spec/quality
  review, plus a final whole-branch review (verdict: ready to merge).

Notes / follow-up

• Detection adapters match sample_id by exact image-file name (flat image dir);
  classification still honours labels.id_strategy. Documented in
  docs/modules/data/configuration.md. A follow-up to unify detection with
  id_strategy (+ e2e YOLO/VOC dispatch tests) is recommended.

Checklist

• CI — Required checks are green
• Breaking changes — Not a breaking change: data.labels.format defaults
  to native, which preserves prior behaviour exactly. Title has no !.
• Contributor guide — I've read Pull requests and commit messages
  before requesting review.

Optional

• Issue (optional) — Closes Implement pluggable label-format adapters (COCO / YOLO / VOC ingestion) #338.
• Docs (optional) — User config option + label-format table; contributor
  "Adding a label format" guide.
• Tests (optional) — Per-adapter unit tests + COCO detection/classification
  end-to-end tests through the dispatch seam (test_label_formats.py).