Ncnn: ShuffleNetV2 performance decreased on X86

[RangiLyu]

Environment:
CPU: Intel(R) Xeon(R) E-2124 CPU @ 3.30GHz 4 Core
System: Windows10 19042.685

version 2020/11/10

loop_count = 4
num_threads = 4
powersave = 0
gpu_device = -1
cooling_down = 1
        mobilenet_v2  min =    7.11  max =   11.80  avg =    8.57
       shufflenet_v2  min =    4.92  max =    5.27  avg =    5.11
     efficientnet_b0  min =   11.14  max =   11.38  avg =   11.21
            resnet18  min =   25.91  max =   27.43  avg =   26.49

version 2020/12/28

loop_count = 4
num_threads = 4
powersave = 0
gpu_device = -1
cooling_down = 1
        mobilenet_v2  min =    6.86  max =    7.85  avg =    7.16
       shufflenet_v2  min =   10.54  max =   11.28  avg =   10.78
     efficientnet_b0  min =   11.13  max =   11.48  avg =   11.28
            resnet18  min =   21.15  max =   21.64  avg =   21.31

新版本下ShuffleNetV2的耗时变成了之前的两倍。

11月下旬时更新了很多x86 pack4的内容，shufflenet相比测试的其他模型多了Concat、Slice和ChannelShuffle，ChannelShuffle没有更新过，Concate 和 Slice都更新了x86 pack4，可能是这两个Layer的问题。

[nihui]

修好了～