Fix multi-collection filtering in search/query/vsearch

[1kuna]

Problem

When multiple -c/--collection flags are passed to qmd search, query, or vsearch, results are often empty or incomplete.

Root cause: With a single -c, the collection filter is pushed into the SQL query (AND d.collection = ?). With multiple -c flags, qmd performs a global unfiltered top-K search then post-filters by qmd://{collection}/ prefix. Large collections dominate the global top-K, so post-filtering removes everything from smaller requested collections.

Fix

Push the collection filter into the underlying retrieval for all three search paths instead of post-filtering a limited result set:

• FTS search: WHERE ... AND d.collection IN (?, ...) via a shared SQL helper
• Vector search: Precompute eligible hash_seq values for requested collections and apply AND hash_seq IN (...) directly in the sqlite-vec KNN query
• Hybrid query: Same filter pushed through to both FTS and vector sub-queries

Added normalizeCollectionFilter() and appendCollectionFilterSql() helpers in store.ts to handle single vs. multi-collection SQL generation consistently.

Type changes

• searchFTS and searchVec (both the Store methods and the exported functions) now accept string | string[] for the collection parameter
• HybridQueryOptions.collection and VectorSearchOptions.collection updated to string | string[]

Tests

Added 4 test files/scenarios (TDD, written before the fix, confirmed failing, then passing):

• test/cli.test.ts — end-to-end CLI regression with 80-doc noisy collection vs 2 small target collections
• test/store.test.ts — searchFTS with multi-collection array filter
• test/query-collection-routing.test.ts — verifies hybridQuery/vectorSearchQuery pass array filters through
• test/searchvec-multi-collection.test.ts — vector search under top-K domination with collection IN filter

Full test suite passes (5 pre-existing LLM timeout failures unrelated to this change).