Add Option for Final Chunk Embedding Model to be different

[jparkerweb]

this can be important when using prefixes. it doesn't make since to use prefix model to create semantic similarity since they need separate document/query prefixes when creating embedding before calculating cosine similarity. but, it could be useful to specify a model that uses prefixes for the final embedding output.

• calculate chunks and combine with a model like L6
• on final chunk return, optionally use a different embedding model that works with "prefixes" (modern-bert, etc).