feat: 模型推理端到端闭环 — 真实流式推理 + 释放/删除 + GPU 状态同步

后端 (platform.py + platform_store.py):
- 新增 _build_messages_payload() 转换前端格式为 OpenAI messages
- 新增 _stream_chat_proxy() SSE 流式代理到算力节点
- 新增 _unload_from_compute_node() 真正释放算力节点 GPU 显存
- 重写 model_compare_load: 从假 PID/端口改为真正调用算力节点加载模型
- 修复 model_compare_unload: 调用 _unload_from_compute_node 释放 GPU
- 修复 model_compare_delete: 先释放 GPU 再删除记录
- 修复 model_compare_stream_chat: 从 mock 改为 StreamingResponse 代理
- 修复 model_chat_local/stream: 消息格式转换 + 路径修正
- PlatformStore 新增 _inference_nodes 追踪，gpus() 同步推理占用状态
- preload/unload 端点标记/清除推理节点占用

算力节点 (compute):
- inference.py: 适配新版 LLaMA-Factory API (get_infer_args 4 返回值、ChatModel args dict、stream_chat 新签名)
- inference.py: unload() 增加 gc.collect + torch.cuda.empty_cache + synchronize 彻底释放显存
- main.py: inference/load 移除 HTTPException(500)，错误以 200 正常返回

前端:
- InferenceChatView: 真实模式下走 SSE 流式推理，mock 模式保留兼容
- InferenceCreateView: 调用 preloadLocalModel + createCompare 真实创建推理任务，失败回退 mock
- InferenceListView: 「停止」改为「释放」，删除前先释放算力节点，改进错误提示
- compare.ts: 新增 streamChatReal() fetch SSE，preload 超时提升至 5 分钟
- useStreamChat.ts: send() 支持 useMock 参数，真实模式调用 streamChatReal
- GPU 选择过滤: 仅显示在线算力节点上的空闲 GPU

Co-Authored-By: Claude <noreply@anthropic.com>

backend/app/db/platform_store.py

@@ -293,6 +293,19 @@ class PlatformStore:
         self.database_url = _psycopg_url(database_url or settings.database_url)
         self.ensure_schema()
         self.ensure_seed_data()
+        # Track which compute nodes have an active inference model loaded
+        self._inference_nodes: set[str] = set()
+
+    # ── inference node tracking ────────────────────────────────────
+
+    def mark_inference_loaded(self, node_id: str) -> None:
+        self._inference_nodes.add(node_id)
+
+    def mark_inference_unloaded(self, node_id: str) -> None:
+        self._inference_nodes.discard(node_id)
+
+    def is_inference_loaded(self, node_id: str) -> bool:
+        return node_id in self._inference_nodes
 
     @contextmanager
     def connect(self) -> Iterator["PgConnection"]:
@@ -2698,6 +2711,11 @@ class PlatformStore:
                 )
                 busy = task is not None and task.get("status") == "running"
                 reserved = task is not None and task.get("status") in {"syncing", "queued"}
+                # Also mark GPU as busy if an inference model is loaded on this node
+                inference_busy = self.is_inference_loaded(row["node_id"])
+                if inference_busy and not busy:
+                    busy = True
+                    reserved = False
                 memory_used = round(row["memory_total_gb"] * (0.72 if busy else 0.18 if reserved else 0.04), 1)
                 gpu_percent = 86 if busy else 22 if reserved else 3
                 memory_total = float(row["memory_total_gb"] or 0)