Replace Harbor-covered benchmark implementations with thin Harbor compatibility wrappers

[neubig]

Context

Several benchmarks in this repository overlap with Harbor registry datasets/adapters. Once Harbor execution is centralized, these benchmarks can be reduced to thin compatibility wrappers that preserve OpenHands CLI/output conventions while delegating task execution to Harbor.

Candidate benchmarks

• SWE-Bench -> swebench-verified
• SWE-Bench Pro -> swebenchpro
• SWE-Bench Multilingual -> swebench_multilingual
• SWTBench -> swtbench-verified
• SWESmith -> swesmith
• Multi-SWE-Bench -> multi-swe-bench
• GAIA -> gaia
• Terminal-Bench -> terminal-bench@2.0
• SWEGym -> swegym

Proposed scope

• Add a small registry/mapping from OpenHands benchmark names to Harbor dataset names.
• Convert covered benchmarks incrementally to call the shared Harbor runner.
• Preserve current script entry points and output.jsonl/cost-report behavior where practical.
• Leave unsupported or parity-sensitive behavior on the legacy path until a Harbor adapter is validated.

Acceptance criteria

• At least one non-Terminal-Bench covered benchmark uses the shared Harbor wrapper path.
• The migration path is documented for the remaining covered benchmarks.
• Tests cover mapping and output conversion behavior.