Batch auto-label a folder with a model + review loop

[wkentaro]

What to build

A batch auto-label + review workflow: run an AI model over a folder of images to pre-populate annotations, then let the user step through and accept/correct them. This builds directly on the per-shape human/AI provenance already in the data model (#2185), which distinguishes AI-proposed shapes from human-confirmed ones.

This is HITL: it needs UX/architecture decisions (how runs are triggered, how review/accept is surfaced, how partial results are persisted).

Acceptance criteria

• A user can run a model across a folder and have proposed annotations saved with AI provenance
• A review flow lets the user accept/edit/reject proposed shapes
• Accepted shapes are marked as human-confirmed via the existing provenance metadata

Blocked by

• AI inference crash-guarding (Guard AI inference run so model errors don't silently crash the app #2202 / When I click on Create AI-polygon the program crashes and flashes back #1300 cluster) should land first so batch runs fail gracefully