fix: 性能优化、前端转写器配置、任务进度丢失及 MLX Whisper 回退问题修复

### 性能优化
- 后端任务执行从串行锁改为 ThreadPoolExecutor 并发执行（默认3线程）
- 添加 GZipMiddleware 响应压缩 + Nginx gzip 配置
- 数据库连接池参数优化（pool_size=10, max_overflow=20）
- 视频帧提取并行化（ThreadPoolExecutor）
- LLM 重试配置缓存到实例，避免每次请求读 env var
- 前端路由级代码拆分（React.lazy + Suspense）
- Vite manualChunks 拆分 markdown/markmap/vendor
- MarkdownViewer 用 React.memo + useMemo 减少不必要渲染
- NoteHistory Fuse.js 实例 useMemo 缓存
- useTaskPolling 无待处理任务时跳过轮询
- 移除 antd 依赖（NoteForm Alert、modelForm Tag），改用 shadcn/ui

### 前端转写器配置（新功能）
- 新增 TranscriberConfigManager（JSON 文件存储，替代环境变量）
- 新增 GET/POST /transcriber_config API 端点
- 新增 GET /transcriber_models_status 模型下载状态查询
- 新增 POST /transcriber_download 后台模型下载触发
- 前端转写器设置页面：引擎选择、模型大小选择、模型下载管理
- deploy_status 端点同步从配置文件读取

### Bug 修复
- 修复任务进行中切换页面后进度丢失：Home.tsx status 派生逻辑补全中间状态
- 修复 MLX Whisper 静默回退 fast-whisper：移除环境变量门控，macOS 下自动尝试导入
- MLX Whisper 不可用时抛出 RuntimeError 而非静默回退
- 前端展示 MLX Whisper 可用性状态，不可用时禁用保存

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

backend/app/utils/video_reader.py

@@ -3,6 +3,7 @@ import hashlib
 import os
 import re
 import subprocess
+from concurrent.futures import ThreadPoolExecutor, as_completed
 import ffmpeg
 from PIL import Image, ImageDraw, ImageFont
 
@@ -54,6 +55,18 @@ class VideoReader:
             return mm * 60 + ss
         return float('inf')
 
+    def _extract_single_frame(self, ts: int) -> str | None:
+        """提取单帧，返回输出路径或 None（失败时）。"""
+        time_label = self.format_time(ts)
+        output_path = os.path.join(self.frame_dir, f"frame_{time_label}.jpg")
+        cmd = ["ffmpeg", "-ss", str(ts), "-i", self.video_path, "-frames:v", "1", "-q:v", "2", "-y", output_path,
+               "-hide_banner", "-loglevel", "error"]
+        try:
+            subprocess.run(cmd, check=True)
+            return output_path
+        except subprocess.CalledProcessError:
+            return None
+
     def extract_frames(self, max_frames=1000) -> list[str]:
 
         try:
@@ -61,14 +74,22 @@ class VideoReader:
             duration = float(ffmpeg.probe(self.video_path)["format"]["duration"])
             timestamps = [i for i in range(0, int(duration), self.frame_interval)][:max_frames]
 
+            # 并行提取帧
+            max_workers = min(os.cpu_count() or 4, 8, len(timestamps))
+            frame_results: dict[int, str | None] = {}
+            with ThreadPoolExecutor(max_workers=max_workers) as pool:
+                futures = {pool.submit(self._extract_single_frame, ts): ts for ts in timestamps}
+                for future in as_completed(futures):
+                    ts = futures[future]
+                    frame_results[ts] = future.result()
+
+            # 按时间戳顺序整理结果，并进行去重
             image_paths = []
             last_hash = None
             for ts in timestamps:
-                time_label = self.format_time(ts)
-                output_path = os.path.join(self.frame_dir, f"frame_{time_label}.jpg")
-                cmd = ["ffmpeg", "-ss", str(ts), "-i", self.video_path, "-frames:v", "1", "-q:v", "2", "-y", output_path,
-                       "-hide_banner", "-loglevel", "error"]
-                subprocess.run(cmd, check=True)
+                output_path = frame_results.get(ts)
+                if not output_path or not os.path.exists(output_path):
+                    continue
 
                 if self.dedupe_enabled:
                     frame_hash = self._calculate_file_md5(output_path)