Port ProgramBench benchmark to Harbor

[neubig]

Context

benchmarks/programbench currently contains substantial benchmark-specific inference and evaluation logic for cleanroom task images and ProgramBench submission/evaluation behavior. A Harbor adapter is needed before this benchmark can be removed from the bespoke OpenHands path.

Proposed scope

• Create a Harbor adapter/dataset for ProgramBench tasks.
• Preserve cleanroom task image usage, workspace submission format, and upstream evaluation semantics.
• Support resource settings appropriate for ProgramBench long rebuild tasks.
• Run parity against the current programbench-infer/programbench-eval workflow.

Acceptance criteria

• ProgramBench tasks can be run through Harbor.
• Evaluation output matches upstream ProgramBench expectations.
• Parity results and resource caveats are documented.