perf: keep-alive upstream sessions, TTL cache on upstream data, /stream LRU, cache hls.min.js

- shared thread-local curl_cffi sessions (no TLS handshake per call)
- in-memory TTL cache: search 5min, details/episodes/languages/
  playlist-URLs 1h; api_stream retries once with a cold cache on 502
  (recovers rotated playlist URLs)
- /stream LRU (24 entries) serves re-seeks and playlist re-fetches from memory
- hls.min.js: Cache-Control public max-age=30d (was no-store -> 500KB
  re-download per page load on mobile)

Measured (click->play chain, cold -> warm):
  search    0.45s -> 0.00s
  detail    0.46s -> 0.00s
  episodes  0.40s -> 0.00s
  stream    1.06s -> 0.00s
  m3u8 x2   0.64s -> 0.00s
  segment   0.20s -> 0.00s

README.md

@@ -47,6 +47,9 @@ the app or CLI are rejected immediately, even with an old cookie.
 - Search with poster grid, detail view with synopsis + season tabs
 - HLS playback (vendored hls.js), browser ABR across 1080/720/360p
 - Dub/sub toggle
+- Fast repeat navigation: upstream data is TTL-cached and HLS
+  playlists/segments are LRU-cached, so re-opening a show or re-seeking is
+  near-instant (first load still pays the upstream round-trips)
 - Autoplays the first episode (or resumes the last watched one); if the
   browser blocks autoplay, a "click to play" overlay appears