fix(server): flush outbound queue before closing ports + silence CI trace spam

[st0o0]

Summary

• CloseAllPorts() data loss fix: When the ApplicationBridge's upstream
  finish triggered CloseAllPorts(), any response data still queued in
  _outboundQueue was silently discarded — the method called
  Complete(_outNetwork) without draining the queue first. Under CI load
  (Linux, Docker matrix variant), this race caused the H1.0 E2E roundtrip
  test to fail with "HTTP/1.0 connection closed before response body was
  complete." Now routes through CompleteAfterFlushingOutbound() when the
  queue is non-empty.

• Tracing leak in ServerStartupReliabilitySpec: The test configured
  global Debug-level Senf tracing on stderr but never disabled it. Since
  Servus.Senf.Tracing is a process-wide singleton, all subsequent tests
  inherited the debug listener — flooding CI logs with thousands of
  [Debug][Stage] lines in the Docker matrix variant. Added
  Tracing.Disable() in the finally block.

Test plan

• H1.0 E2E roundtrip tests pass (previously flaky on Linux CI)
• Full E2E suite passes (101 tests, verified 5× locally)
• Full unit suite passes (5961 tests)
• Docker CI variant no longer floods logs with Debug trace output