fix: cancel open request timeouts when the server connection stage stops

[pjfanning]

Motivation

RequestTimeoutSupport schedules a request's timeout through materializer.scheduleOnce and keeps the TimeoutAccess in openTimeouts until the response for that request comes back, which is when the timeout is cancelled.

The stage had no postStop, so a connection that goes away with requests still open — the client disconnecting, the connection being aborted, or the server being torn down — left one scheduled task per open request behind. Those tasks keep the request, the timeout handler and the stage's async callback reachable for as long as the configured request-timeout (20 seconds by default), and firing them achieves nothing: they only invoke a callback on a stage that is already gone.

Modification

Cancel everything still in openTimeouts in postStop, using the same best-effort clear() the regular response path uses.

Result

Connections that end before their requests are answered no longer leave scheduled timeouts behind.

Tests

• New test in HttpServerWithExplicitSchedulerSpec: RequestTimeoutSupport is driven with probes, a request gets a timeout handler that records being called, and both sides are then completed with the request still unanswered. With the explicitly triggered scheduler the test can then let the timeout elapse and assert the handler never runs. Reverting only the postStop addition makes it fail.
• sbt "http-core/testOnly org.apache.pekko.http.impl.engine.server.HttpServerSpec org.apache.pekko.http.impl.engine.server.HttpServerWithExplicitSchedulerSpec" — pass (66 tests), including the five existing request-timeout tests.
• sbt "http-core/mimaReportBinaryIssues" — pass.
• scalafmt --mode diff-ref=upstream/main --test — pass.

References

None — found while auditing src/main for resource leaks.