feat: findEmbeddingsByIds read primitive

[techiejd]

Adds `findEmbeddingsByIds` (public) / `findByIds` (adapter contract): fetch stored embedding records by primary key — including the raw `embedding` vector, which the normal search/query API never returns. Building block for "more like this."

The `id` of each record is whatever `search()` returns as `result.id`, so a search result round-trips directly.

What's here

• `EmbeddingRecord` type + `VectorizedPayload.findEmbeddingsByIds({ knowledgePool, ids })` + `DbAdapter.findByIds(payload, poolName, ids)` signatures, re-exported for adapters.
• Public wiring in `createVectorizedPayloadObject` (empty `ids` short-circuits to `[]`), covered via the in-memory mock adapter.
• PostgreSQL: Drizzle `inArray` direct read; surfaces the raw pgvector column as `number[]`; non-numeric ids dropped via `/^\d+$/`.
• MongoDB: `find({ _id: { $in } })`; non-24-hex ids dropped via the existing `HEX24` guard (never throws).
• Cloudflare: official `binding.getByIds(ids)` (typed by refactor(cf): adopt @cloudflare/workers-types #61); metadata + values mapped to the record, no mapping-collection lookup needed.
• Docs: `adapters/README.md` contract (interface, method table, `EmbeddingRecord` types reference, custom-adapter stub) + root `README.md` Local API section.

Contract

Misses dropped (result length may be `< ids.length`); order not guaranteed; empty `ids` → `[]` without a backend call; unknown/malformed ids treated as misses, not errors.

Tests

All green: main `test:int` 86/86 · pg 65/65 · cf 93/93 · mongodb 100/100. Each adapter has a dedicated `findByIds` spec (full record incl. numeric embedding + reserved + extension fields, drop-misses, empty-ids).

Stacking

Based on #61 (`feat/cf-workers-types`), which the CF `findByIds` depends on (unlocks the typed `getByIds`). Merge #61 first, then this retargets to `main`.