SubpixelConv2d is too slow when construct a graph

[ShieLian]

The reason may be in layers.py#L2191, the loop adds too many ops.
An alternative way is as follow:
Assume input feature maps X is b*h*w*(r^2*c) tensor,
Xs=tf.split(X,r,3) #b*h*w*r*r
Xr=tf.concat(Xs,2) #b*h*(r*w)*r
X=tf.reshape(Xr,(b,r*h,r*w,c)) # b*(r*h)*(r*w)*c
Then X is the return value in L2197

This way only add three ops in the graph, and for an feature map 242464, the time constructing the graph is much more short.
Also, time for constructing train-ops is shorten.

layers.py#L2188 to L2196 need to change, and L2170 to L2186 can be removed.

[zsdonghao]

Thank you for your advice. As it should be your credit, It is possible that you can contribute your code via push request?
Many thanks.

[ShieLian]

Create PR #176