feat(client): localize card art to the UI language - #7070

Changing the UI language now swaps displayed card art to the same
printing in that language. The printing the user chose is preserved --
only its image is exchanged for its localized sibling -- and cards with
no sibling in that language keep their English art.

The localized printing id comes from MTGJSON `foreignData[].identifiers
.scryfallId`, which we already download for the text sidecars, so this
needs no new upstream dependency. Scryfall image URLs are constructible
from the id alone, and the English id is already embedded in the URLs we
store, so the sidecar is a bare id->id map rather than a second copy of
the image data. English users download nothing extra.

The generator walks AllSetFiles one set at a time rather than parsing
AllPrintings whole: a whole-file jq parse peaks at ~3.5 GB RSS against
~95 MB per-set, a 36x reduction.

Localization is applied in `localFaceImageUrl`, the single funnel every
stored-URL path reaches, and composes with `splitSizedImageUrl` so the
card back and Scryfall's "image coming soon" placeholder come back
byte-identical -- `isPlaceholderImageUrl` compares with `===`, so
rewriting them would silently disable the printing-fallback chain.

`imageRequestCache` keys on the art vocabulary a URL was produced with
(`de` vs `de:pending`) rather than the bare language, so the locale
map's arrival is a genuine key change; the printing-selection caches
stay language-neutral because which printing wins is a function of art
preferences, not language.

Known limitation: deck-builder search results hold URLs from search
time, so switching language mid-search leaves them English until the
next search.

Six findings from the PR review, all confirmed against the code first:

- Workbox never routed the locale maps in production. Workbox's
  RegExpRoute refuses a cross-origin match that does not begin at index 0
  of the href, and these are served from R2 — so the bare suffix pattern
  silently skipped the very requests the rule exists for. Now two
  anchored branches, mirroring the engine-WASM rule.

- CI re-downloaded the ~169 MB AllSetFiles archive every run: the
  MTGJSON cache is saved before the generator runs and GitHub caches are
  immutable. Adds a restore/save pair for the five generated maps,
  mirroring the draft-pools idiom — caching the small output rather than
  adding ~1 GB of extracted set files to a shared entry. The generator
  already no-ops when the maps exist, so a hit skips the download.

- deploy-cf.sh ran every upload in its own background subshell, each
  doing a read-modify-write through one shared "$DEPLOY_CACHE.tmp": two
  workers interleaving lose entries. Workers now record a private tag
  file and a single writer merges after the wait loop. Pre-existing, but
  five more entries made it materially more likely.

- The locale-art load ran during render, where it wrote module-global
  state that decides whose fetch may install. A discarded concurrent
  render could let an uncommitted language win that race; it now runs in
  a [language] effect placed after the invalidation subscription.

- PrintingPickerModal resolved tile URLs during render with no
  subscription, so opening it while the map was in flight left English
  art with no re-render on arrival. Adds `useLocaleArt`, which owns the
  load and the invalidation tick. useCardImage deliberately does not use
  it — it has an oracleId-filtered subscription, and an unfiltered second
  one per tile would resurrect the re-render storm that filter prevents.

- Adds the missing request-deduplication test, and a picker test that
  drives a mounted modal through pending -> resolved.