fix: preserve startup errors and exit status

[LauraGPT]

Summary

• log unhandled startup exceptions with their traceback
• preserve a nonzero process exit status when initialization fails
• retain the existing immediate-exit behavior used after shutdown
• add a regression test for startup failures

Root cause

The top-level entrypoint currently calls os._exit(0) from an unconditional finally block. Any exception raised while loading handlers is therefore discarded before Python can report it, and Docker sees a successful exit even though the service never started.

This matches the symptom in #296: handler registration succeeds, loading stops partway through initialization, no traceback is shown, and the container exits as status 0. The reported log only proves that SenseVoice was registered; its model load had not started yet. The next handler after the last successful load message is Silero VAD.

This change keeps the explicit os._exit behavior but logs unexpected exceptions and propagates exit status 1, so the underlying handler error becomes actionable.

Verification

• python3 -m pytest -q tests/test_demo_entrypoint.py (1 passed)
• python3 -m py_compile src/demo.py tests/test_demo_entrypoint.py
• git diff --check

Related to #296.