fix: route session/update racing with a reaped holder to the global handler (IJAI-2705)

[EugeneTheDev]

Summary

• A session/update can find a speculative session holder just before withInitializingSession reaps it, then queue into it afterwards. handleOrQueue awaited the failed session and threw AcpExpectedError: Session … not found, which Protocol logged at SEVERE.
• handleOrQueue now reports such an update as undelivered, and the handler sends it to globalSessionUpdateHandler (or drops it at debug level).
• failAndDrainQueue closes the channel before draining, so an update can't be queued after the drain and lost. The holder fails only after the drain, so late updates come after the queued ones.
• ClientSessionHolder moved to its own internal class to be testable. No public API change.

Tests

• ClientSessionHolderTest drives both states deterministically with runTest. Both tests fail on the old code with the ticket's error.

YouTrack: IJAI-2705