feat(lancedb): implement InsertDocuments for LanceDbVectorIndex

[abhicris]

Summary

Implements the InsertDocuments trait for LanceDbVectorIndex, matching the existing implementations for MongoDB / Postgres / Qdrant / Milvus / ScyllaDB / SurrealDB / HelixDB / S3 Vectors / SQLite. LanceDB was one of only two integrations still missing this trait (Neo4j is the other, and is covered by #1636).

Before this PR, users had to hand-build Arrow RecordBatches to insert data into LanceDB; they can now do:

let index = LanceDbVectorIndex::new(table, model, "id", SearchParams::default()).await?;
let docs = EmbeddingsBuilder::new(model).documents(my_docs)?.build().await?;
index.insert_documents(docs).await?;

Behaviour

• Each Embedding produces one row. Document fields are flattened onto that row (or stored under a document field when the serialized doc is not a JSON object), and the embedding vector goes into the table's FixedSizeList<Float32|Float64> column.
• The embedding column is taken from SearchParams::column when set; otherwise it is auto-detected as the sole FixedSizeList<Float32|Float64> column in the table schema (the same default LanceDB uses for search column inference). An explicit error is returned when zero or multiple candidate columns exist.
• Supports both Float32 and Float64 embedding columns; embedding dimensionality is validated against the schema before the write is attempted.
• Uses arrow_json::ReaderBuilder::build_decoder().serialize(&rows) to decode serialized docs against the non-embedding portion of the schema, then splices the embedding column back in to produce a batch that matches the full table schema. This avoids having to reimplement per-column Arrow conversion for every scalar type.

Implementation notes

• Added arrow-json = "56" as a dependency (aligned with the workspace's arrow-array = "56").
• No public API change to LanceDbVectorIndex / SearchParams; the trait impl reads the embedding column name from existing config.
• RecordBatchIterator + Table::add(...).execute() is used for the write (same API pattern the examples already use for initial table creation).

Test plan

Added insert_documents_round_trip to tests/integration_tests.rs:

• Seeds a LanceDB table with one row using the existing fixture schema (id, definition, embedding as FixedSizeList<Float64, 1536>).

• Calls InsertDocuments::insert_documents with two fresh (Word, Embeddings) pairs driven by a mocked OpenAI embeddings endpoint.

• Asserts that Table::count_rows() grows from 1 to 3 after the insert.

• Asserts that the two new ids (inserted-1, inserted-2) are returned by top_n via the same VectorStoreIndex handle — proving that the rows were written to the correct columns and the embedding column was populated with a valid FSL vector.

• New integration test added and covers the happy path.

• cargo check -p rig-lancedb / cargo test -p rig-lancedb --test integration_tests — unable to run locally (host disk full); happy to rebase / adjust on CI feedback.

---

— kcolbchain / Abhishek Krishna