Send pub/sub messages on the server's event loop

Problem

PubSub.publish() is called from the StationWatcher thread and from test output callbacks (StationServer.publish_final_state), not from the tornado event loop that owns the websocket connections. sockjs-tornado's websocket transport calls write_message() directly, so off-loop:

• the frame is written to the IOStream from the wrong thread, which isn't thread-safe, and
• write_message() schedules its completion task on the calling thread's event loop. StationWatcher creates that loop but never runs it, so every task stays pending for the life of the process.

At shutdown every one of these tasks gets reported:

Task was destroyed but it is pending!
task: <Task pending name='Task-8' coro=<WebSocketProtocol13.write_message.<locals>.wrapper() running at .../tornado/websocket.py:1111>>

A station that runs for a shift with the GUI open logs thousands of these lines on exit.

Fix

PubSub records the serving loop when the first client subscribes (on_open always runs on that loop). publish() sends directly when it's already on that loop, and otherwise hands the send to the loop with call_soon_threadsafe. If nobody has subscribed, it does nothing, as before. If the loop has closed, the publish is dropped, because the server is shutting down. _loop is stored per subclass, like _lock and subscribers, so StationPubSub and DashboardPubSub don't share it.

Testing

• New test/output/servers/pub_sub_test.py: publishing from another thread, from the serving loop, with a client filter, with no subscribers, and after the loop has closed. Three of its cases fail without the fix.

• Full suite passes: pytest test (588 passed).

• I also ran a script that starts a StationServer, connects a websocket client and runs 5 tests that each log heavily and then prompt. The client answers each prompt over HTTP.

  	"Task was destroyed" at exit	Prompt latency (median)	Plug POSTs
  master	8–10	0.22–0.39 s	5 × 200
  this PR	0	0.20–0.27 s	5 × 200

🤖 Generated with Claude Code