feat(extensions): add LatticeWasmEmbeddings provider

[ohdearquant]

Summary

Adds LatticeWasmEmbeddings, a local embeddings provider for ruvector-extensions backed by @khive-ai/lattice-embed-wasm, a pure-Rust BERT-family text embedder compiled to WebAssembly (source: ohdearquant/lattice/npm/lattice-embed-wasm). It plugs into the existing EmbeddingProvider seam next to the OpenAI, Cohere, Anthropic, and HuggingFace providers, giving a native-Rust local-embedding option that does not depend on @xenova/transformers. Offered as a convenience alternative at parity with the existing HuggingFaceEmbeddings local path, not a faster or higher-quality replacement.

What it does

• getMaxBatchSize() returns 1. The wasm package has no batch API (one embed(text, model) call per text), so this reports that honestly instead of simulating a larger batch.
• getDimension() returns 384 for both supported models (minilm, bge-small); the constructor throws immediately for any other model name.
• embedTexts() dynamic-imports the package (mirroring how HuggingFaceEmbeddings imports @xenova/transformers), with an install-hint error if missing. A null result from the wasm layer (unknown model, unsupported over wasm, or unresolvable weights) is a real thrown error for an explicitly-selected provider, never a silent skip.
• Output is L2-normalized and symmetric (no query/passage prefixing at the wasm layer).
• @khive-ai/lattice-embed-wasm is an optional peer dependency (peerDependenciesMeta.optional), matching the openai/cohere convention. Not a hard dependency.

Test plan

• Construction throws for an unknown model name
• getDimension() returns 384 for minilm (default) and bge-small
• getMaxBatchSize() returns 1
• Live embed test (real package + real weights): asserts exact 384-length output and L2 norm ~1.0; self-skips when the optional package or its weights are unavailable (e.g. a fresh CI checkout with no local model cache).

Note (limitation of the underlying package, not this PR): the wasm package's release-asset weight-fetch tier is not live yet, so a user with no local model cache and no model-dir override currently gets a thrown error on first embed. Local-cache and env-override paths work today.