perf: improved speed of BDI

By efficiently checking when the next work has to be done, we can
increase performance not having to "busy loop". Time from transcription
-> message to LLM agent is now down to sub 1 millisecond.

ref: N25B-316

src/control_backend/agents/perception/vad_agent.py

@@ -63,7 +63,7 @@ class VADAgent(BaseAgent):
 
         self.audio_buffer = np.array([], dtype=np.float32)
         self.i_since_speech = settings.behaviour_settings.vad_initial_since_speech
-        self._ready = False
+        self._ready = asyncio.Event()
         self.model = None
 
     async def setup(self):
@@ -141,14 +141,11 @@ class VADAgent(BaseAgent):
         while await self.audio_in_poller.poll(1) is not None:
             discarded += 1
         self.logger.info(f"Discarded {discarded} audio packets before starting.")
-        self._ready = True
+        self._ready.set()
 
     async def _streaming_loop(self):
+        await self._ready.wait()
         while self._running:
-            if not self._ready:
-                await asyncio.sleep(0.1)
-                continue
-
             assert self.audio_in_poller is not None
             data = await self.audio_in_poller.poll()
             if data is None: