Allow external usage proxy concurrency

[bingran-you]

Summary

• multiplex external usage-proxy routes over one fixed local listener using secret path prefixes
• allow high concurrency for external usage tracking in a single BenchFlow process while still rejecting multi-worker fixed-port sharing
• persist rollout cleanup failures into result.json so downstream experiment health gates cannot treat failed teardown as a reusable healthy slot
• update docs and add regression coverage for concurrent Daytona-style external proxy usage and cleanup telemetry

Tests

• uv run ruff check .
• uv run ty check src
• uv run python -m pytest tests/test_usage_tracking.py tests/test_trajectory_proxy_path_prefix.py tests/test_usage_proxy.py tests/test_acp.py::TestSandboxStartupDiagnostics -q
• uv run python -m pytest tests/ -q

Notes

• Follow-up to PR Add Daytona usage tracking proxy support #568: fixed-port external usage tracking can now support concurrent rollouts in one process; sharded workers still need separate ports/tunnels.
• Kept as draft because the requested live Daytona E2E for deepseek-v4-flash is blocked by the current provider proxy model list: no healthy deepseek-v4-flash model ID is exposed.

---

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.