refactor: the native downloaders share one download loop, file record and seen-ledger key; a clean walk marks its credential verified

Patreon, SubscribeStar and Discord each carried their own copy of the same
per-post download loop, file record and seen-ledger key, and every copy was
another place a fix could be missed.

- BaseNativeDownloader runs one post's files: the seen and on-disk skips,
  the mid-post time-box, pacing, validation and the per-file sidecar. Each
  platform overrides only where its layout differs: Discord's channel folder,
  file names and message-id sidecar, and Patreon's Mux video through yt-dlp.
- One MediaItem and one ledger_key in native_ingest_common replace the three
  per-client copies. Keys are unchanged, except that a SubscribeStar file with
  no gallery id is now keyed by its filename instead of an empty id.
- DownloadResult, ErrorType and the run-stats and tier-gated helpers move
  from gallery_dl.py, which now serves only HentaiFoundry, to
  download_result.py. The unused delegator methods on GalleryDLService go.
- The download lock's platform list is the native-ingester list, not a copy.
- A walk that used a credential and finished with no error type stamps that
  credential's last-verified time, so the credentials page shows when the
  account last worked. A tier-limited walk does not count: an expired session
  hides paid posts the same way.
- The extension no longer clears a retired pixiv token on startup.
- Two comments that still described every download as gallery-dl now name
  both backends.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

backend/app/celery_app.py

@@ -4,7 +4,7 @@ Queues:
   import      — filesystem import path (FC-2)
   ml          — WD14 + SigLIP inference (FC-2; runs in the ml-worker image)
   thumbnail   — image and video thumbnail generation (FC-2)
-  download    — gallery-dl tasks (FC-3)
+  download    — source walks: the native ingesters and gallery-dl (FC-3)
   scan        — periodic source checks (FC-3) — kept separate so long imports
                 don't starve the scheduler
   maintenance — recovery sweeps, pHash backfill, GPU-queue coordination, etc.