refine mkldnn benchmark script

[tensor-tang]

Avoid cpu conflicting when use_mkldnn=False

[luotao1]

避免CPU竞争后，在Intel(R) Xeon(R) CPU E5-2620 0 @ 2.00GHz 上，对train vgg 64 False（使用MKLML）测得的对比数据如下：

• 优化前的脚本：一个batch需要72s。

I0927 17:22:12.967727 11210 TrainerInternal.cpp:165]  Batch=1 samples=64 AvgCost=6.90691 CurrentCost=6.90691 Eval:  CurrentEval: 

I0927 17:23:24.291543 11210 TrainerInternal.cpp:165]  Batch=2 samples=128 AvgCost=6.89256 CurrentCost=6.8782 Eval:  CurrentEval: 

I0927 17:24:35.678067 11210 TrainerInternal.cpp:165]  Batch=3 samples=192 AvgCost=6.86233 CurrentCost=6.80187 Eval:  CurrentEval: 

I0927 17:25:46.317230 11210 TrainerInternal.cpp:165]  Batch=4 samples=256 AvgCost=6.80972 CurrentCost=6.65189 Eval:  CurrentEval:

• 优化后的脚本：一个batch仅需30s

I0927 17:30:13.134626 31875 TrainerInternal.cpp:165]  Batch=1 samples=64 AvgCost=6.97192 CurrentCost=6.97192 Eval:  CurrentEval: 

I0927 17:30:43.715212 31875 TrainerInternal.cpp:165]  Batch=2 samples=128 AvgCost=6.92188 CurrentCost=6.87183 Eval:  CurrentEval: 

I0927 17:31:15.349555 31875 TrainerInternal.cpp:165]  Batch=3 samples=192 AvgCost=6.86882 CurrentCost=6.7627 Eval:  CurrentEval: 

I0927 17:31:45.824867 31875 TrainerInternal.cpp:165]  Batch=4 samples=256 AvgCost=6.81989 CurrentCost=6.67311 Eval:  CurrentEval: 

[luotao1]

效果很棒。但建议脚本进行简化，现在脚本的分支太多了。

[tensor-tang]

Done,

[luotao1]

LGTM