Add voice state tags, SuperTTS configs, and demo WS log groups.

Parse leading <state> tags from LLM replies and emit response.state over the product websocket while stripping tags from TTS/text streams. Add FastGPT+Xfyun voice configs (including state-enabled preset), SuperTTS support, and context sync for interrupted turns. Refresh the voice demo with a state indicator and collapsible audio delta websocket log groups.

Co-authored-by: Cursor <cursoragent@cursor.com>

config/voice.json

@@ -47,7 +47,13 @@
   "agent": {
     "system_prompt": "You are a helpful, friendly voice assistant. Keep responses concise and natural for spoken conversation.",
     "greeting": "Please introduce yourself briefly.",
-    "greeting_mode": "generated"
+    "greeting_mode": "generated",
+    "response_state": {
+      "enabled": false,
+      "tag": "state",
+      "event_type": "response.state",
+      "max_prefix_chars": 256
+    }
   },
   "services": {
     "stt": {