feat: migrate AI_VoiceOverData_Vanilla MP3s to Git LFS

Voice data (9,555 MP3s, ~1.2 GiB) was tracked as plain git blobs,
bloating .git to 1.1 GiB per clone. Migrated to LFS so clones drop to
single-digit MiB; the actual MP3 bytes live once on Gitea LFS storage
(/srv/gitea/data/lfs on git03), fetched on demand via 'git lfs pull'.

Also marks *.mp3 export-ignore so 'git archive' (and the release
workflow's build_zip.sh) skips them — release zips will contain the
addon framework + lookup tables only, never the binary data. Voice
files distributed separately (TBD).

.gitattributes

@@ -1 +1,2 @@
 * -text
+*.mp3 filter=lfs diff=lfs merge=lfs -text export-ignore