Need to optimize the memory allocation of VGG16 model in Fluid

[kuke]

Maximum feasible batch size on TITAN X (Pascal, 12GB global memory)

	cifar10 ( 3 x 32 x 32 )	flowers (3 x 224 x 224)
Fluid	966	19
TensorFlow	1618	33

• CMAKE_BUILD_TYPE:STRING=Release
• CUDNN: v6

Compared to TensorFLow, the VGG16 model in Fluid can only run with much smaller batch size right now.

[chengduoZH]

This issue's mission has accomplished.