CUDA: fix race condition in FA vector kernels (#13742)

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

@@ -212,6 +212,7 @@ static __global__ void flash_attn_vec_ext_f16(
                 }
             }
             if (__all_sync(0xFFFFFFFF, skip)) {
+                __syncthreads();
                 continue;
             }
 #endif // GGML_USE_HIP

ggml/src/ggml-cuda/fattn-vec-f32.cuh

@@ -217,6 +217,7 @@ static __global__ void flash_attn_vec_ext_f32(
                 }
             }
             if (__all_sync(0xFFFFFFFF, skip)) {
+                __syncthreads();
                 continue;
             }
 #endif // GGML_USE_HIP