fix(910c/glm52): start_vllm_docker.sh 加per-TP参数覆盖 + 修printf转义破坏JSON
问题: add16实际启动走start_vllm_docker.sh(用config.env全局MAX_MODEL_LEN=131072),
而非engine_build_server_args(有per-TP覆盖但只写server_cmd.txt不影响启动).
TP=8用131072上下文+无per-TP覆盖,且engine_build_server_args的printf %q转义
破坏了JSON(--additional-config的enable_dsa_cp等未生效),致MoE tiling失败.
修复:
1. start_vllm_docker.sh: 加per-TP case覆盖(TP=8:16384/64/0.95, TP=16:131072/256/0.92)
2. run_adaptive_concurrency_add16.sh: engine_build_server_args 用单引号包裹替代
printf %q,避免JSON被反斜杠转义(仅影响server_cmd.txt记录,实际启动走start_vllm_docker.sh)