fix(BIZ-42): critical deadlock + major review issues

- Fix #1 (Critical): evaluate_retreat() deadlock — changed _retreat_lock to
  threading.RLock() to allow reentrant acquisition when evaluate_retreat()
  calls get_429_rate() while holding the lock
- Fix #2 (Major): Dashboard queue chart now uses snap.queue.per_priority
  instead of Math.random() mock data
- Fix #3 (Major): structlog uses JSONRenderer instead of ConsoleRenderer
  for JSON-format output as required by acceptance criteria
- Bonus: webui.py _build_snapshot() now async and includes queue data
  with per-priority depth for dashboard consumption

Reviewed-by: 梁思筑 (architect)
Co-authored-by: multica-agent <github@multica.ai>

services/nvidia_sidecar/rate_limiter.py

@@ -295,8 +295,8 @@ class AdaptiveTokenBucket(TokenBucket):
         # 上次状态变更时间
         self._last_state_change: float = time.monotonic()
 
-        # 避退状态锁
-        self._retreat_lock: threading.Lock = threading.Lock()
+        # 避退状态锁（RLock 防止 evaluate_retreat() → get_429_rate() 重入死锁）
+        self._retreat_lock: threading.RLock = threading.RLock()
 
     # ---- 429 反馈 ----