feat(embedding): built-in local backend with on-demand model download

[mrsibe]

What does this PR do?

Implements the built-in local embedding backend (#52) and its download transport (#53).

With no remote embedding connection configured, KnowNote now indexes and searches
with a bundled runtime and an on-demand-downloaded model, instead of failing
with No embedding model configured. The first time a user actually uses the
knowledge base (upload / create knowledge base / search), the model is downloaded
once to userData/models and then works fully offline. A remote connection, when
configured, still takes precedence and keeps the existing AI SDK path.

Why?

The RAG path required the user to understand protocol / base URL / API key /
model id / dimensions before uploading a first PDF. Public issues (#33, #35, #31,
#14, #10, #12) all trace back to this. Embedding should be infrastructure, not a
configuration step: configure chat -> upload a PDF.

Issue #53 covers making the weights arrive reliably: the default downloader in
transformers.js does not verify, resume, or retry, and huggingface.co is not
always reachable.

Related issue

Fixes #52
Fixes #53
Fixes #33
Fixes #37

#37's report is brief, but upload is what triggers indexing, so the zero-config
local default plus the mirror fallback should remove the failure it describes.

Already closed before this PR, listed as context only (no closing keyword):
#35, #31, #14, #10, #12. The dimension-mismatch family (#14 / #31) was fixed by
the per-notebook vector table (#47); this PR adds the zero-config local default
and the EmbeddingSpace identity on top of it.

What changed?

• EmbeddingBackend abstraction: EmbeddingService only says "I want an
  embedding"; LocalEmbeddingBackend (ONNX) and RemoteEmbeddingBackend
  (existing ModelClient) implement it. Remote is used when configured,
  otherwise local.
• Local backend: pinned Xenova/multilingual-e5-small (q8, pinned revision,
  384-dim), E5 query: / passage: prefixes applied by purpose, truncation
  at 512 tokens, normalize. Inference sets env.allowRemoteModels = false.
• Index identity: EmbeddingSpace (model + revision + dtype + pooling +
  normalize + prefixes + dimensions) persisted per notebook in a new
  notebook_embedding_spaces table. Search rejects vectors from a different
  space; the indexing path rebuilds/clears and marks documents pending.
• Download transport: checked-in manifest (sizes + SHA256), self-written
  downloader with SHA256 verification, atomic rename, HTTP Range resume,
  per-file retry/backoff, disk-space pre-check, classified actionable errors,
  source reachability probing (huggingface.co then hf-mirror.com), and
  manual import from a local folder. Download progress is broadcast to the UI.
• Settings UI: Embedding section with Built-in Local / Custom provider
  and Not installed / Downloading / Installed states, model path, size, delete,
  import and source probing.
• Packaging: @huggingface/transformers, onnxruntime-node and sharp are
  production dependencies with asarUnpack entries; the packaged smoke test
  loads them.

How was this tested?

• npm run typecheck (node + web + test) — passes.
• npm run lint — 0 errors.
• npm test — 49 tests pass (22 pre-existing + 27 new).
• npx electron-vite build — main/preload/renderer bundle builds; the
  @huggingface/transformers dynamic import stays external as intended.
• Real end-to-end run in Node against the real mirror: downloaded all 7
  manifest files (140,461,908 bytes, every SHA256 verified), then loaded the
  pinned model with allowRemoteModels = false and embedded text offline:
  dimensions === 384, all components finite, L2 norm 1.0000.
• Semantic sanity check: a Chinese query scored 0.904 against its matching
  Chinese document vs 0.804/0.670 for others, and an English query scored 0.870
  against the Eiffel Tower document — confirming the query/document prefixes and
  cross-lingual behavior.
• Tests cover space identity, manifest/buildResolveUrl, the downloader (resume,
  checksum mismatch cleanup, retry, cancel, probing, error classification),
  model registry/aliases and local-backend prefixes.

Not run: npm run smoke:packaged / build:unpack (needs the full electron-builder
matrix locally). The packaged smoke check was extended, but should be exercised in CI.

Notes / deferred

• transformers.js v4.3.0 does not forward the pinned revision during its file
  discovery step, which breaks offline (allowRemoteModels=false) loading.
  This PR works around it by hardlinking the manifest files at the revision-less
  cache path; the revision directory stays canonical so a revision bump still
  forces a fresh download.
• A user-configurable custom download base URL is not exposed in the UI yet;
  huggingface.co -> hf-mirror.com probing plus manual import is shipped. It
  is a small follow-up if enterprise/internal mirrors are needed.

Screenshots / recordings

UI change: the Embedding section in Settings now offers Built-in Local vs Custom
provider with install state. No screenshot captured in this environment.

Checklist

• I have reviewed my own changes.
• npm run typecheck passes.
• npm run build passes (electron-vite build).
• I have tested the affected user workflow (real model download + inference).
• I have not included unrelated changes.
• I have updated documentation when necessary.

Desktop / build changes

• Not applicable
• npm run build:unpack passes.
• npm run smoke:packaged passes.