Fix TopicManager shutdown deadlock with connected clients

[jonyMarino]

Problem

Closing Runtime or Unreal with a topic client still connected can hang indefinitely. TopicManager::Stop holds manager_lock_ while nng_close waits for the disconnect callback, and that callback calls UnsubscribeAll, which needs the same lock.

Change

Release the topic-data lock while joining the receiver and closing the socket. A separate lifecycle mutex serializes complete Start/Stop transitions, and repeated calls are harmless. Reacquire the data lock before resetting socket and callback state.

Add bounded subprocess regressions covering connected clients, pending receive frames, concurrent Stop calls, and repeated Start/Stop cycles. A watchdog fails the child after 10 seconds instead of hanging the test runner.

Validation

• Windows Release suite: 374/374 passed.
• New regressions: 20 repetitions each, all 40 passed.
• Negative control: connected-client regression fails after 10 seconds with the original TopicManager implementation.
• Runtime and Unreal 5.7 editor-game both reported commit 27a1825c1a0bfdf1500e0e65be51cba33e3fd2bf and exited normally with clients/publications active: Runtime 0.409 s, Unreal 1.956 s. Each completed 5 scene reloads and 3 reconnects first; Unreal also delivered valid RGB/depth. These focused shutdown checks did not repeat flight commands. No test processes/listeners remained.