perf(worker): 按实例内存上限扩展异步容量

将策略推导出的集群目标与单 Worker 内存安全上限分离，避免失活实例的全部容量转移到单个 2 GiB Pod。

生产环境单实例上限和媒体物化并发设为 24；两实例正常时总容量由 16 提升至 48，后续可通过增加 Worker 节点继续扩展。新增 0093 迁移保存实例容量上限，并兼容滚动发布中的旧实例。

风险：更高媒体并发会增加 Worker 和节点内存压力，保留单实例 24 的硬边界并由持续监控观察 RSS、OOM、队列与节点余量。

验证：go test ./... -count=1；go vet ./...；真实 PostgreSQL 分配与故障转移测试；迁移安全检查；kubectl kustomize；gofmt -l；git diff --check。

apps/api/internal/config/config.go

@@ -63,6 +63,7 @@ type Config struct {
 	MediaMaterializationConcurrency               int
 	AsyncQueueWorkerEnabled                       bool
 	AsyncWorkerHardLimit                          int
+	AsyncWorkerInstanceHardLimit                  int
 	AsyncWorkerRefreshIntervalSeconds             int
 }
 
@@ -123,6 +124,7 @@ func Load() Config {
 		MediaMaterializationConcurrency:   envIntValidated("AI_GATEWAY_MEDIA_MATERIALIZATION_CONCURRENCY", 8),
 		AsyncQueueWorkerEnabled:           env("AI_GATEWAY_ASYNC_QUEUE_WORKER_ENABLED", "true") == "true",
 		AsyncWorkerHardLimit:              envIntValidated("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT", 2048),
+		AsyncWorkerInstanceHardLimit:      envIntValidated("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT", 32),
 		AsyncWorkerRefreshIntervalSeconds: envIntValidated("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS", 5),
 	}
 }
@@ -150,6 +152,9 @@ func (c Config) Validate() error {
 	if c.AsyncWorkerHardLimit < 1 || c.AsyncWorkerHardLimit > 10000 {
 		return errors.New("AI_GATEWAY_ASYNC_WORKER_HARD_LIMIT must be between 1 and 10000")
 	}
+	if c.AsyncWorkerInstanceHardLimit < 1 || c.AsyncWorkerInstanceHardLimit > 10000 {
+		return errors.New("AI_GATEWAY_ASYNC_WORKER_INSTANCE_HARD_LIMIT must be between 1 and 10000")
+	}
 	if c.AsyncWorkerRefreshIntervalSeconds < 1 {
 		return errors.New("AI_GATEWAY_ASYNC_WORKER_REFRESH_INTERVAL_SECONDS must be positive")
 	}