Switch embed to llama_get_embeddings_seq

[iamlemec]

Due to updates in ggml-org/llama.cpp#5796, sequence level embeddings are now output through a separate channel from token level embeddings, and they are accessed with llama_get_embeddings_seq.

[abetlen]

@iamlemec thank you! Just so I understand, the sequence level embeddings are the ones that are pooled up to the end of the from the last processed batch?

Also, I think the new function in llama_cpp.py is duplicated by accident.

[iamlemec]

Oh yeah, I didn't see you added it already! Yup, it's the pooled embeddings by sequence for the last batch. So for both mean pooling and cls (first token) it works, and for non-pooling it returns null.