feat(gateway): worker 池主备分组派发，主全挂才用备机

- pool.py: WorkerState 增加 primary 标记；acquire_worker 分级派发——
  主组有在线只在主组内挑空闲（全忙排队，不退备组），主组全部下线才用备机
- config.py: 新增可选 primary_workers 配置（须为 workers 子集，自动过滤+告警）；
  不配置/空列表 = 全部视为主 worker，兼容旧行为
- config.json(不入 git): 3 台 GPU worker 设为主，max_global_concurrency 1→3
- app.py: /health 增加 workers_primary_total / workers_primary_healthy
- tests: 新增 4 条主备派发单测（主优先/主全挂用备/主全忙不走备/无配置兼容）
- docs(实现说明): 补充 primary_workers 配置说明，改配置需重启网关生效

gateway/config.py

@@ -15,6 +15,7 @@ CONFIG_PATH = Path(__file__).resolve().parent / "config.json"
 DEFAULTS = {
     "public_base_url": "https://hair.xiangsilian.com",
     "static_dir": "static/annotations",
+    "primary_workers": [],  # 主 worker 地址列表（须为 workers 的子集）；空 = 全部视为主 worker
     "health_check": {
         "path": "/health",
         "interval_seconds": 8,
@@ -83,6 +84,15 @@ def load_config() -> dict:
         if not isinstance(w, str) or not w:
             raise ValueError(f"配置错误: workers[{i}] 必须是非空字符串")
 
+    primary_workers: List[str] = cfg.get("primary_workers", [])
+    for i, w in enumerate(primary_workers):
+        if not isinstance(w, str) or not w:
+            raise ValueError(f"配置错误: primary_workers[{i}] 必须是非空字符串")
+    for w in primary_workers:
+        if w not in workers:
+            logger.warning("⚠ primary_workers 中的 %s 不在 workers 列表里，将被忽略", w)
+    cfg["primary_workers"] = [w for w in primary_workers if w in workers]
+
     password: str = cfg.get("shared_password", "")
     if not password or password == "REPLACE_ME_ROTATE_PERIODICALLY":
         logger.warning(