[RNE Rewrite] Add a small standalone on-device vector store

Summary

Follow-up to the embeddings work in #1292. Add a small on-device vector store with similarity search as a standalone feature — not coupled to the embedding flow.

Motivation

#1292 adds text/image embedders that return L2-normalized vectors, but consuming them still requires the app to keep its own array of vectors and hand-roll cosine ranking (as both the nlp and computer-vision demos currently do). A minimal, reusable vector store would cover the common "embed once, query many" use case (semantic search, RAG retrieval, zero-shot ranking) without every app reimplementing it.

Proposal

A small, self-contained utility that:

• Stores a set of vectors with associated metadata/ids.
• Supports add / remove / query by nearest neighbors (cosine / dot-product), returning top-k with scores.
• Is decoupled from the embedders — it takes plain vectors as input, so any source (our embedders or external) works.

Kept intentionally minimal (in-memory to start; persistence can be a later increment).

Explicitly out of scope

• Not part of the embedding pipeline/API — it lands as its own feature in a separate PR.
• No external vector-DB dependency for the first pass.

Context

• Raised by @barhanc in review of #1292: a small vector DB should be its own separate functionality, not folded into the embedding flow.
• Depends on / follows: #1292.

Start by reviewing the embedding work in #1292 and the existing vector-ranking logic in the nlp and computer-vision demos. Define the standalone in-memory store around plain vectors, ids, and metadata, with add, remove, and top-k similarity queries using cosine or dot-product scores. Done means the utility is independent of the embedders and supports the listed operations without an external vector database.