* feat: detect the spoken language for models without native detection
* feat: download the language detector with the models that need it
* feat: add language detection settings and show the real language status
* feat: record each transcription's language in history
* minor review notes
* tooltip fix, better auto info
* tweak some ui for the model
* move to experimental and download
* review nits
* log change
* feat: use shared model catalog
* fix: preserve other model quants on deletion
* fix: offer retired model replacements
* fix: preserve unreadable model caches
* fix: report model deletion failures
* refactor: simplify model availability and cache reuse
* fix: match cached repos to the catalog by exact name
HF repo ids are case-insensitive, so another tool can cache a catalog repo
under other casing. Matching it case-insensitively relabelled the file with
the catalog's casing, which on a case-sensitive filesystem points at a
folder that doesn't exist, so the model vanished. Exact matching (as on
main) keeps such a file listed under the casing it was found at.
* fix: harden model deletion and catalog check on Windows