CUDA/HIP: fix race in flash_attn_ext_f16_process_tile when nbatch_combine != DKQ/2 (#30103)

Suggested-by: Johannes Gäßler <johannesg@5d6.de>

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1772,7 +1772,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 }
             }
         }
-        if (np > 1) {
+        if (np > 1 || nbatch_combine != DKQ/2) {
             __syncthreads();
         }
     }