fix: #1661 Preserve realtime session voice settings when updating agents

[seratch]

Summary

Resolves #1661

• merge run and initial realtime session settings when constructing a session
• reuse the merged base when recomputing settings so the active voice is preserved on handoffs or agent updates
• add a regression test that ensures the initial voice and output audio format remain intact when refreshing session settings

Testing

• make format
• make lint
• make mypy (fails: optional dependencies numpy, sqlalchemy, and litellm are not installed in the test environment)
• make tests (fails: optional dependencies numpy and litellm are not installed in the test environment)