fix(realtime): omit null audio formats in SIP call attach updates

[weiguangli-io]

Summary

Fixes #2308 by avoiding format: null in session.update payloads when attaching to an existing SIP call via call_id.

In call-attach mode, input_audio_format / output_audio_format can be intentionally omitted. The SDK was still serializing these as explicit null values, which can cause the Realtime API to reject the update with errors like:

• Invalid type for 'session.audio.input.format': expected an object, but got null

What changed

• In OpenAIRealtimeWebSocketModel._get_session_config:
  • only set audio.input.format if a non-null format can be converted.
  • only set audio.output.format if a non-null format can be converted.
• Added regression tests to verify:
  • with call_id and no explicit formats, update payload omits both format fields.
  • with explicit formats, payload still includes normalized audio/pcmu fields.

Validation

• uv run --with ruff ruff check src/agents/realtime/openai_realtime.py tests/realtime/test_realtime_model_settings.py
• uv run --with mypy mypy src/agents/realtime/openai_realtime.py tests/realtime/test_realtime_model_settings.py
• env -u ALL_PROXY -u all_proxy -u HTTPS_PROXY -u https_proxy -u HTTP_PROXY -u http_proxy -u NO_PROXY -u no_proxy uv run --with pytest pytest -q tests/realtime/test_realtime_model_settings.py
• env -u ALL_PROXY -u all_proxy -u HTTPS_PROXY -u https_proxy -u HTTP_PROXY -u http_proxy -u NO_PROXY -u no_proxy uv run --with coverage --with pytest coverage run -m pytest -q tests/realtime/test_realtime_model_settings.py

Changed executable lines coverage (local):

• src/agents/realtime/openai_realtime.py: 100% (6/6)