Fix OpenHands SDK Modal startup and full-eval reliability

[AlienKevin]

Summary

• preserve OpenHands -oh image entrypoint in the SDK adapter so agent-server starts correctly
• avoid repo import shadowing in OpenHands sandboxes by using / workdir
• make Modal eval backend use explicit keepalive startup (entrypoint([]) + sleep infinity) for stable sb.exec
• update vendored OpenHands SDK/tools code for Python 3.11 compatibility

Why

• fixes repeated Modal Sandbox ... not found failures during run/eval
• enables end-to-end reproduction of gemini-3-flash + openhands_sdk on full CooperBench

Scope

• src/cooperbench/agents/openhands_agent_sdk/adapter.py
• src/cooperbench/eval/backends/modal.py
• vendored OpenHands SDK/tools compatibility fixes under src/cooperbench/agents/openhands_agent_sdk/

Validation

Reproduced Gemini 3 Flash + OpenHands performance on CooperBench:

{
  "pass_rate": 0.2846153846153846,
  "total_runs": 652,
  "passed": 185,
  "failed": 465,
  "errors": 2,
  "skipped": 0,
}