Add inpainting evaluator, dataset schema, and canonical metric

Problem

winml eval does not register the canonical inpainting task, although runtime inference can use an image + binary mask contract.

Both eval --schema --task inpainting and artifact evaluation reject the task before dataset loading. This prevents task-metric evidence for direct-ONNX inpainting models such as opencv/inpainting_lama.

Requested capability

• Register an inpainting evaluator and schema.
• Define the required image, mask, and ground-truth restored-target columns.
• Document a canonical metric or metric set for inpainting quality.
• Keep input preprocessing aligned with the runtime image + mask contract.

Reproduction

winml eval --schema --task inpainting

Observed with WinML CLI 0.2.0:

Task 'inpainting' is not supported by winml eval.

A public image+mask dataset can exercise input shape, but a valid evaluator also needs authoritative restored targets and a documented metric; scores must not be inferred from an input-only dataset.