feat(search)!: batch facet searches into a single multi_search

[ddeboer]

Fix #532

The keyed DatasetFacets surface resolved each selected facet through its own engine.search call, so a typical listing page fanned out into 1 listing search + N facet searches. This PR batches the whole facet selection into one engine round-trip, restoring the pre-migration single-multi_search behaviour: a typical page now costs the listing search plus one batched facet search (with the label cache on, exactly 2 Typesense round-trips). Recorded as ADR 5.

Port (@lde/search) — breaking

• SearchEngine gains the batch entry point searchFacets(searchType, queries), returning one FacetsOutcome per query, positionally: { facets } or { error } (an allSettled-style union). A failure of one query is reported in place and never discards its siblings’ facets; the promise rejects only for batch-level failures (foreign type, invalid query, transport). Engines answer every query facet-only — as if limit: 0 and without ordering, whatever the query carries — so hits are never transferred. Existing SearchEngine implementations must add the method.
• The executable port-contract suite (@lde/search/testing) covers the new method: foreign-type rejection, per-query validation, positional alignment (including a non-zero-limit query), and the empty batch.

Typesense adapter (@lde/search-typesense)

• searchFacets compiles each query with the existing buildSearchParams (normalized to per_page: 0, no sort_by) and sends the whole batch as one multi_search.
• A failed multi_search entry (reported inline by Typesense) becomes a per-query error outcome naming the affected facet fields, so its siblings’ facets survive; fetchLabels now also throws on an inline error entry, engaging the label degradation path (onLabelError, id-only references) instead of silently missing every label.
• Reference-facet labels for the entire batch resolve in one bundled lookup, or from the in-memory cache when labelCacheTtlMs is set; the label-degradation logic is shared between search and searchFacets.

GraphQL surface (@lde/search-api-graphql)

• Selected facet fields load through a per-request DataLoader (facet-batch.ts): same-tick loads are collected, grouped by their effective where after skip-own-filter — facets whose own field is unfiltered share one query, so the unfiltered browse collapses to a single facet query; each own-filtered facet gets its own variant — and dispatched as one engine.searchFacets call.
• Facet counts, skip-own-filter semantics, per-facet failure isolation (onFacetError fires only for the fields of the query that actually failed), reference labels, and the public SDL are unchanged. Facet-only queries drop the listing orderBy (nothing to order).
• New direct dependency: dataloader (GraphQL Foundation’s reference batching implementation) instead of a hand-rolled microtask batcher.

Deliberately not done

• Riding same-where facets on the main listing search via facet_by for a 1-round-trip browse — needs GraphQLResolveInfo selection-walking; inside one multi_search the difference is negligible (see ADR 5).
• Overlapping the listing search with the facet dispatch (pre-existing sequencing; the facet queries derive only from the resolved query, so this is a possible follow-up).