fix: 分布式其他卡要提前同步，等 0 卡跑完

Signed-off-by: YdrMaster <[email protected]>

Author: YdrMaster

models/llama/cuda/src/nccl_parallel.rs

@@ -172,6 +172,8 @@ fn test_infer() {
                                 );
 
                                 next.send(pair.idx() as _).unwrap()
+                            } else {
+                                stream.synchronize();
                             }
                         }
                     });