fix: reuse event loop across Harbor attempts

[nmogil]

Summary

Fix the real multi-attempt Daytona activation failure found by the approved 18-attempt provider-parity matrix.

The first attempt completed successfully, but the runner created and closed a fresh event loop around every HarborBackend.run_attempt. Daytona SDK 0.196.0 shares an aiohttp client session tied to the first loop, so attempts 2–18 failed before sandbox creation with Event loop is closed.

This change:

• Executes every prepared Harbor attempt sequentially within one owned asyncio.run(...) lifecycle.
• Preserves attempt ordering and existing single-attempt behavior.
• Adds a regression test proving multiple attempts observe the same live loop and that the owned loop closes after the run.
• Adds no concurrency, retries, provider bypasses, or evidence/cleanup weakening.

Live failure boundary

• Attempt 1: quality eligible.
• Attempts 2–18: failed before sandbox creation/model execution.
• Aggregate output correctly failed closed at 1/18 evidence.
• Nothing was uploaded to BlindBench.
• Provider check after failure: zero remaining Daytona sandboxes.

Verification

Supervisor-run from the final diff:

• Focused event-loop/matrix tests: 2 passed
• Full non-live suite including real Docker smoke: 115 passed, 1 paid-live deselected
• Ruff: passed
• mypy: passed (15 source files)
• git diff --check: passed

No paid calls were made while implementing this fix.