* refactor bias tensor variable names * use create_tensor_qkv for jina-bert-v2
* model: using single llm_build per arch * fix merge * nits