Resumable saves: ranged downloads that survive dropped connections, reloads and app kills

Server
- /api/download/:id answers Range with a strong ETag ("<id>.<gen>") and honours
  If-Range; a stale partial gets the whole current file (streamed, so Bun does
  not re-apply the Range itself). Uploads get the same treatment.
- GET /api/download/:id/prepare never blocks: ready {gen,size,sha256,etag,ext},
  working (server still fetching), legacy (HEVC / too long / cache offline),
  failed. Recently refused prepares are remembered for 10 minutes.

Browser
- The OPFS worker saves in 8 MiB ranges, writes at the byte offset, flushes
  each chunk, retries each chunk 6 times with backoff (30 s idle timeout) and
  keeps the .part plus a .part.json sidecar naming the server copy it belongs
  to. A changed copy restarts cleanly; the finished file is hashed once and
  checked against the server's SHA-256.
- SaveQueue remembers unfinished saves and resumes them on start, online,
  return to the foreground and every 2 minutes while visible; one at a time.
- Downloads shows live MB progress, 'Preparing on server', 'Verifying', and
  paused saves with Resume and Cancel (confirmed).
- listVideos ignores the sidecars; new listPartials/discardPartial helpers.

Verified in Chromium through a connection-dropping proxy: paused at 8 MiB,
auto-resumed after a reload from byte 8388608, final SHA-256 matched.

frontend/sw.js

@@ -70,6 +70,7 @@ const SHELL = [
   '/lyrics-core.js',
   '/stats-core.js',
   '/sha256.js',
+  '/resume-core.js',
   '/sha256-wasm.js',
   '/device-db.js',
   '/hash-worker.js',