Fix LlamaIndexEmbeddingOperator returning vector=None for every chunk

[kaxil]

Fixes #68416.

LlamaIndexEmbeddingOperator.execute() documents chunks[*].vector as the embedding for each chunk, but every chunk came back with vector: None. The operator relied on VectorStoreIndex(nodes, ...) setting .embedding on the nodes it was given as a side effect -- but _get_node_with_embedding() attaches embeddings to model_copy() copies, never the originals (same behavior from v0.10 through v0.14.22, so no version pin avoids it). The vectors only ever existed inside the index's vector store.

The fix embeds the original nodes explicitly before any index work:

• embed_model.get_text_embedding_batch() over node.get_content(metadata_mode=MetadataMode.EMBED) -- the exact content embed_nodes() embeds (includes node metadata, respects excluded_embed_metadata_keys), so the vectors are identical to what VectorStoreIndex produced internally before.
• VectorStoreIndex is now built only when persist_dir is set. embed_nodes() skips nodes whose .embedding is already set, so persisting reuses the vectors instead of re-calling the embedding API (pinned by a test that counts embed calls through the real llama-index persist path). Without persist_dir, the index was built and immediately discarded.

Design rationale

Pre-embedding beats the alternative of reading vectors back from index.vector_store.data.embedding_dict: that couples the operator to SimpleVectorStore internals (.data doesn't exist on other stores) and still builds a throwaway index when not persisting. Pre-embedding only uses the public BaseEmbedding API.

The duck-type check in _resolve_embed_model now requires get_text_embedding_batch instead of get_text_embedding, since that's what execute() calls. It's a concrete method on BaseEmbedding (subclasses override _get_text_embeddings, not the public method), so every real embedding class passes the check across the supported llama-index range.

closes #68488
closes #68434

[kaxil]

@vojay-dev can you try with the fix in this PR, please?

[vojay-dev]

@vojay-dev can you try with the fix in this PR, please?

@kaxil

Tested against the deps where I hit the bug (provider 0.4.0, llama-index-core 0.14.22, Python 3.13, Runtime 3.2-5) with MockEmbedding. On 0.4.0 chunks[].vector comes back None (with and without persist_dir); with this PR the vectors are populated directly. Also confirmed no double-embedding: 3 docs with persist_dir embed 3 texts total, not 6.

LGTM, thanks for the fast turnaround 👍!