[6.x] Avoid loading the whole container listing in Asset::exists()

[edalzell]

Stacked on #15610; merge that first. Until it lands this diff includes its commits.

Asset::exists() answered via container()->files(), which unserializes the whole cached recursive listing. Asset::find() calls it, so resolving one asset reference on the front end loaded the full listing. On a 197k-file S3 container that is an 80 MB cache entry and OOMs a 1 GB worker.

exists() now uses the in-memory listing only when this request has already loaded it (AssetContainerContents::isLoaded()), and otherwise does a single disk()->exists() call (one S3 HEAD). CP listings load the listing first, so they keep their zero-extra-IO behaviour. The disk answer is memoized per asset path, since meta() calls exists() before its own memo; without that a front-end find() plus augmenting width/height/size/alt was 16 HEADs per asset, now 3.

I skipped the suggested meta-cache shortcut: move() never forgets the old path's meta key, so it would report a moved-away path as existing. A disk check is always right.

Closes #15607