Add deconvolution operator, and the corresponding implement using cudnn_v3

[tornadomeet]

1. have done a simple deconvolution_shape_check;
2. when using deconvolution layer, the mxnet's implemention and cudnn's implemention give the same output almostly. using the follow example: replace one line of the example/mnist/lenet.py , refer to deconv_lenet.py

[antinucleon]

Thanks for the great PR! Could you fix the lint?

[antinucleon]

You can try make lint to see the lint result locally.

[tornadomeet]

fix lint now.

[antinucleon]

Thanks! Another issue is could you add namespace for enum? for example https://github.com/dmlc/mxnet/blob/master/src/operator/convolution-inl.h#L24 . Because we will soon provide amalgamation version code, this part will make name conflict..

[tornadomeet]

i'll add these and check now.

[tornadomeet]

now have added the namespace for deconvolution-inl.h and cudnn_deconvolution-inl.h.

[mli]

thanks for the PR! but what does deconvolution do? in my understanding it can be used to reserve a convolution layer and used for visualization. but in your lenet example, you just replace the conv layer by deconv layer for training?

[tornadomeet]

deconvolution mostly used for upsampling, for example, in image segmentation, can refer fcn-xs, deconvnet, crfasrnn.
the lenet example here is just for unitest, and i'm implenmenting the fcn-xs in mxnet now, but not very soon.

[mli]

@tornadomeet very helpful, thanks!

[antinucleon]

Last thing, could you add your name in Contributor.md? then I will merge it

[tornadomeet]

ok, i have added my name now.