feat: file-based ONNX loading and asymmetric retrieval prefix API for UserDefinedEmbeddingModel

[CrispStrobe]

Summary

Stacks on #237 — please merge that first; this PR's diff against main will shrink once it does.

This PR extends UserDefinedEmbeddingModel with two ergonomic features frequently requested for large or externally-stored models.

OnnxSource — file-based model loading

Replaces the single onnx_file: Vec<u8> field with an OnnxSource enum:

pub enum OnnxSource {
    Memory(Vec<u8>),   // existing behaviour
    File(PathBuf),     // new: ORT resolves .onnx.data companion automatically
}

• UserDefinedEmbeddingModel::from_file(path, tokenizer_files) — convenience constructor
• UserDefinedEmbeddingModel::new(onnx_source, tokenizer_files) — accepts impl Into<OnnxSource> (both Vec<u8> and PathBuf convert automatically, so existing code is unaffected)
• When OnnxSource::File is used, ORT's commit_from_file is called — this means the runtime resolves any companion .onnx.data file automatically, avoiding the need to load multi-gigabyte weights into RAM

OnnxSource is also unified across the embedding and reranking modules (previously reranking::init had its own identical copy).

Asymmetric retrieval prefix API

Retrieval models often require different prefixes for queries vs documents (e.g. Jina v5: "Query: " / "Document: "):

let model_def = UserDefinedEmbeddingModel::from_file(path, tok)
    .with_query_prefix("Query: ")
    .with_doc_prefix("Document: ");

// query path — prepends query_prefix
model.embed_query(queries, None)?;

// document path — prepends doc_prefix  
model.embed(documents, None)?;

// raw path — no prefix at all
model.embed_raw(texts, None)?;

Test coverage

• tests/local_models.rs: test_jina_v5_nano exercises OnnxSource::File + output key selection + prefix API end-to-end
• tests/text-embeddings.rs: test_user_defined_embedding_model and test_user_defined_reranking_model pass with the new API

🤖 Generated with Claude Code