Port SWE-Bench Multimodal benchmark to Harbor

Context

benchmarks/swebenchmultimodal does not appear to have a direct Harbor registry dataset/adapter yet. Harbor supports multimodal trajectories, but this benchmark needs a task adapter that preserves SWE-Bench Multimodal task setup, resolved-instance selection, and evaluation behavior.

Proposed scope

• Create a Harbor adapter/dataset for SWE-Bench Multimodal.
• Preserve the current resolved-instance selection behavior and multimodal assets.
• Validate the verifier against the current swebenchmultimodal-infer/swebenchmultimodal-eval workflow.
• Document any differences in image handling or multimodal output/trajectory handling.

Acceptance criteria

• SWE-Bench Multimodal tasks can be run through Harbor.
• Parity results are recorded against the existing OpenHands benchmark workflow.
• The OpenHands wrapper can delegate SWE-Bench Multimodal execution to Harbor.

Start by inspecting benchmarks/swebenchmultimodal and the existing swebenchmultimodal-infer and swebenchmultimodal-eval workflow. Compare its task setup, resolved-instance selection, image handling, and trajectory behavior with Harbor and the OpenHands wrapper. Done means tasks run through Harbor, parity results are recorded, and the wrapper delegates execution successfully.