feat: hwaccel + Fanart/TheTVDB + AI + Discover + NFO + Tasks + Recycle bin

Backend
  - service/transcoder.go: encoder selector (software / nvenc / qsv /
    vaapi) with proper hwaccel + scale_* filters per encoder; configurable
    bitrate/preset/height/segment-seconds via transcoder.* config; new
    Active() snapshot + ActiveJob struct for the Tasks panel; unit-tested
    via buildFFmpegArgs(...).
  - service/thetvdb.go: TheTVDB v4 provider — login() caches the JWT for
    24h, SearchSeries() returns Match struct.
  - service/fanart.go: Fanart.tv provider — high-res movie artwork keyed
    by TMDb id; used to upgrade poster/backdrop after a successful match.
  - service/scraper.go: provider chain reworked — anime → Bangumi, tv →
    TheTVDB, default → TMDb, with optional Fanart upgrade post-match.
  - service/discover.go: TMDb /trending/movie/day + /movie/popular for
    the Discover rail.
  - service/ai.go: OpenAI-compatible client. SmartSearch() turns a free-
    form query into a structured SearchIntent JSON; Recommend() emits a
    short list of titles given the user's history. Disabled when
    ai.api_key is empty.
  - service/nfo.go: per-movie .nfo writer (Kodi/Jellyfin schema) +
    library-scope batch exporter.
  - service/media.go: SoftDelete / Restore / Purge / ListRecycleBin
    helpers backed by gorm Unscoped().
  - service/service.go: container wires Fanart, TheTVDB, Discover, AI,
    NFO; everything still tears down cleanly via Close().
  - config: new transcoder.* section (encoder, preset, video_bitrate,
    max_rate, buf_size, max_height, segment_seconds) with sensible
    defaults.

Handlers / routes
  - new files: discover.go, recycle.go, nfo.go, ai.go, tasks.go.
  - registers GET /tasks, /discover/{trending,popular},
    /ai/{status,search,recommend}; admin-only GET /recycle and
    DELETE/POST /media/:id (soft delete / restore / purge); admin-only
    POST /media/:id/nfo and /libraries/:id/nfo.

Frontend
  - new pages: DiscoverPage, TasksPage, RecycleBinPage; SearchPage gains
    an optional AI smart-search toggle that calls /api/ai/search.
  - MediaDetailPage gains admin actions for 导出 NFO and 移至回收站.
  - new api/ helpers: ai.ts, discover.ts, recycle.ts, tasks.ts.
  - App.tsx routes /discover, /tasks, /recycle (last two admin-only).
  - Layout sidebar groups now include Discover + Tasks + Recycle Bin.

Docker / config
  - Dockerfile: adds intel-media-driver / libva-utils / mesa-va-gallium
    so QSV + VAAPI work out of the box on Intel iGPUs.
  - docker-compose.yml: documents devices + group_add + nvidia runtime
    overrides for hardware transcoding.

Verified: go build, go vet, go test (incl. new TestBuildFFmpegArgs across
software / nvenc / qsv / vaapi encoder profiles); tsc -b && vite build
emits 22 route chunks plus the deferred hls chunk; main bundle 248 KB /
83 KB gzipped.

docker-compose.yml

@@ -34,15 +34,27 @@ services:
       MEDIASTATION_DATABASE_DB_PATH: /data/mediastation.db
       MEDIASTATION_CACHE_CACHE_DIR: /cache
       MEDIASTATION_LOGGING_LEVEL: info
+      # MEDIASTATION_TRANSCODER_ENCODER: vaapi    # qsv | vaapi | nvenc | (empty=software)
+      # MEDIASTATION_TRANSCODER_MAX_HEIGHT: 1080
       # MEDIASTATION_SECRETS_JWT_SECRET: change-me-to-a-strong-random-string
       # MEDIASTATION_APP_CORS_ORIGINS: '["https://media.example.com"]'
       PUID: "1000"
       PGID: "1000"
       TZ: Asia/Shanghai
 
-    # Hardware-accelerated transcoding (uncomment when you have an Intel iGPU).
+    # Hardware-accelerated transcoding.
+    #   * Intel iGPU (QSV / VAAPI): mount /dev/dri.
+    #   * NVIDIA NVENC: switch the runtime to nvidia and add `gpus: all`.
     # devices:
     #   - /dev/dri:/dev/dri
+    # group_add:
+    #   - "989"   # match render GID on host
+    # runtime: nvidia
+    # deploy:
+    #   resources:
+    #     reservations:
+    #       devices:
+    #         - capabilities: [gpu]
 
     healthcheck:
       test: ["CMD-SHELL", "wget -qO- http://127.0.0.1:8080/api/health || exit 1"]