Add on_start support to VoiceWorkflowBase and VoicePipeline (#922)

I added an optional `on_start()` method to `VoiceWorkflowBase`, allowing
voice workflows to emit an initial TTS message before any user input.
This aligns with the usage pattern suggested in #488

The `VoicePipeline._run_multi_turn()` method was updated to call this
hook and stream any yielded messages via TTS before transcription
begins. No changes are required to existing workflows, as the method
defaults to returning an empty `AsyncIterator[str]`.

Resolves #488 and thus improves UX for voice agents that need to greet
or instruct users proactively. All tests pass under Python 3.12. Let me
know if you'd like an example workflow or docs added in a follow-up!

Author: vrtnis

src/agents/voice/pipeline.py

@@ -125,6 +125,12 @@ async def _run_multi_turn(self, audio_input: StreamedAudioInput) -> StreamedAudi
                 self._get_tts_model(), self.config.tts_settings, self.config
             )
 
+            try:
+                async for intro_text in self.workflow.on_start():
+                    await output._add_text(intro_text)
+            except Exception as e:
+                logger.warning(f"on_start() failed: {e}")
+
             transcription_session = await self._get_stt_model().create_session(
                 audio_input,
                 self.config.stt_settings,

src/agents/voice/workflow.py

@@ -32,6 +32,14 @@ def run(self, transcription: str) -> AsyncIterator[str]:
         """
         pass
 
+    async def on_start(self) -> AsyncIterator[str]:
+        """
+        Optional method that runs before any user input is received. Can be used
+        to deliver a greeting or instruction via TTS. Defaults to doing nothing.
+        """
+        return
+        yield
+
 
 class VoiceWorkflowHelper:
     @classmethod