fix: avoid unsupported native local embedding batch mode

Legacy issue: investigate true llama-cpp native multi-sequence batch support for local embedding models such as bge-small-zh-v1.5-f16 (current runtime reports n_seq_max=1, so embed_batch uses sequential mode).

Co-authored-by: GPT-5.4 <noreply@openai.com>

Author: Mijamind719