fix: reliable speaking indicator via direct store writes + polling safety net

Eliminate double-state architecture (useState → useEffect bridge → store)
that lost speaking events due to React 18 batching. ActiveSpeakersChanged
now writes speakingParticipantIds directly to voiceStore; 200ms poll
catches missed SDK events. Each VoiceUser subscribes to its own identity
via fine-grained selector for minimal re-renders. Also adds connection
quality indicator and ConnectionInfoPopover.

packages/web/src/components/voice/GlobalAudioRenderer.tsx

@@ -74,9 +74,10 @@ export function GlobalAudioRenderer() {
   const watchingStreams = useVoiceStore((s) => s.watchingStreams);
   const streamAttenuationEnabled = useVoiceStore((s) => s.streamAttenuationEnabled);
   const streamAttenuationStrength = useVoiceStore((s) => s.streamAttenuationStrength);
+  const speakingParticipantIds = useVoiceStore((s) => s.speakingParticipantIds);
 
   // Determine if someone is currently speaking (for stream attenuation)
-  const someoneIsSpeaking = participants.some((p) => !p.isLocal && p.isSpeaking);
+  const someoneIsSpeaking = participants.some((p) => !p.isLocal && speakingParticipantIds.has(p.identity));
 
   // Only render audio for remote participants
   const remoteParticipants = participants.filter((p) => !p.isLocal);