[NFC] polish colossalai/kernel/cuda_native/csrc/kernels/softmax_kernels.cu code style (#954)

colossalai/kernel/cuda_native/csrc/kernels/softmax_kernels.cu

+#include <cooperative_groups.h>
 #include <math.h>
 
 #include <cub/block/block_load.cuh>
@@ -6,8 +7,6 @@
 #include "block_reduce.h"
 #include "kernels.h"
 
-#include <cooperative_groups.h>
-
 namespace cg = cooperative_groups;
 const float EPSILON = 1e-8f;
 
@@ -304,8 +303,7 @@ __global__ void ker_attn_softmax_bw(T *grad, const T *inp, int softmax_length) {
   cg::thread_block b = cg::this_thread_block();
   cg::thread_block_tile<WARP_SIZE> g = cg::tiled_partition<WARP_SIZE>(b);
 
-  for (int i = 1; i < WARP_SIZE; i <<= 1)
-    sum += g.shfl_xor(sum, i);
+  for (int i = 1; i < WARP_SIZE; i <<= 1) sum += g.shfl_xor(sum, i);
 
 #pragma unroll
   for (int i = 0; i < ITERATIONS; ++i) {