Classify the particles in und.toml by language and kind #749

Part of #747. Phase 1 converted the particles mechanically into und.toml under the generic particle kind (111 words on master d3730550; recompute with uv run python tools/vocab/query.py --kind particle --count).

The work: move each particle into the language files where it occurs and give it its kind: family particle, patronymic marker (bin, ibn), article (al-), clan, and so on. A word may sit in several files: a language file means the words found in names written in that language (decision 4 in decisions/vocabulary-model), so de belongs to several. may_be_given is already stated on every particle and moves with it.

Also: per-word particle facts in decisions/ prose move into the entries (note, issue, excluded, counts), and the prose becomes a query pointer, as the suffix pilot did.

Why first: the top candidate packs (Malay, the Spanish and Portuguese surname span, Arabic) all depend on particles, and the multi-language particles are the hardest test of decision 4.

Done when und.toml holds no particle, the compiled sets are unchanged and the snapshot does not move.