fix: skip redundant name_embedding computation in create_entity_node_embeddings

[GraphiteEdgeR]

Summary

create_entity_node_embeddings() unconditionally computes embeddings for all nodes with a non-empty name, even if they already have a valid name_embedding. This PR adds a simple filter to skip nodes that already have their embedding set, making the function idempotent and avoiding redundant API calls.

Problem

During add_episode(), resolved nodes (merged to existing graph entities) go through create_entity_node_embeddings() which calls embedder.create_batch() for all nodes. Since get_entity_node_return_query() deliberately excludes name_embedding from its return fields (to reduce query payload), resolved nodes arrive with name_embedding=None and get re-embedded unnecessarily.

However, if callers pre-load the embedding (via node.load_name_embedding() or node_load_embeddings_bulk()), the current code still re-computes it because there's no check for existing values.

Change

`diff

• filter out falsey values from nodes

• filtered_nodes = [node for node in nodes if node.name]

• Only compute embeddings for nodes that need them (have a name but no existing embedding)

• filtered_nodes = [node for node in nodes if node.name and node.name_embedding is None]
  `

Benefits

1. Idempotent: Safe to call multiple times without redundant API calls
2. Enables optimization: Callers can now pre-load embeddings from DB (via load_name_embedding() / node_load_embeddings_bulk()) before calling this function, and the pre-loaded values will be respected
3. Zero risk: Nodes without embeddings still get computed as before

Suggested follow-up

For maximum benefit, a follow-up PR could add a pre-loading step in extract_attributes_from_nodes() before calling create_entity_node_embeddings():

`python

Pre-load existing name_embeddings for resolved nodes

nodes_missing = [n for n in nodes if n.name and n.name_embedding is None]
if nodes_missing:
await clients.driver.graph_operations_interface.node_load_embeddings_bulk(
clients.driver, nodes_missing
)
`

This would eliminate redundant embedding API calls for all resolved nodes (typically 2-4 per episode).