test: reduce CI flakiness in Ray-backed and smoke tests

[Copilot]

Summary

Tighten test reliability for failures seen on main: intermittent Ray worker/actor startup crashes in unit tests and transient external HTTP failures in the tutorial smoke test. The changes target test isolation and failure handling without weakening the underlying assertions.

Changes

• ZMQ test fixture isolation

  • Replace per-test os.environ mutation for PLUGBOARD_FLAGS_ZMQ_PUBSUB_PROXY with scoped DI settings overrides.
  • Apply the change consistently across shared, unit, and integration ZMQ fixtures to avoid cross-test/process races when Ray workers start in parallel.

• Targeted reruns for Ray-backed flakes

  • Add limited reruns to the Ray-backed unit tests that intermittently fail during actor/worker startup:
    • test_multiprocessing_channel
    • test_state_backend_init
    • test_state_backend_init_with_existing_job
  • Keep reruns narrow to preserve signal and avoid broad pipeline inflation.

• Smoke test handling for external network instability

  • Retry tutorial execution once on transient network failures.
  • If the failure remains clearly external (ConnectError, ReadTimeout, RemoteProtocolError, etc.), skip instead of failing the pipeline.
  • Preserve normal failures as hard failures.

testing_settings = Settings.model_validate({"flags": {"zmq_pubsub_proxy": zmq_pubsub_proxy}})
with override_settings(testing_settings):
    yield ZMQConnector