fix(ipc): drop stale ring on failed re-registration

[WilliamK112]

Summary

• remove the stale connector ring entry before closing the old ring during re-registration
• leave the connector without a registered ring if reopening the new shared-memory segment fails, instead of keeping a closed ring in _ring_registry
• add a regression test for the failed-reopen path that avoids the Windows-incompatible IPC socket fixture

Fixes #209

Validation

• Rebased onto current NVIDIA/xr-ai@ec4169aa3e3c8a57f1d7f59d3e2d97ad9163a8bb
• Confirmed IPC hub: connector re-register left bound to closed ring if ShmRingBuffer reopen fails #209 is open/unassigned with no comments, and open PR searches for 209, _ring_registry, and failed re-registration only return this PR
• uv run pytest -v test_participant_events.py::test_connector_reregistration_open_failure_drops_stale_ring (from tests/) - 1 passed
• uv run python -m py_compile ..\server-runtime\xr_media_hub\ipc\_hub.py test_participant_events.py (from tests/)
• uv tool run --from ruff==0.15.16 ruff check server-runtime\xr_media_hub\ipc\_hub.py tests\test_participant_events.py
• git diff --check origin/main..HEAD

Local environment note: full uv run pytest -v test_participant_events.py on this Windows shell reports the new #209 regression test passing, while the existing four fixture-backed tests error before their test bodies because this local libzmq build does not support the ipc://... transport used by conftest.py.