[Operator] add nearest neighboor up sampling

[antinucleon]

No description provided.

[antinucleon]

@winstywang @piiswrong

[piiswrong]

Looks good to me.
BTW, I also need multi input upsampling (basically upsampling + concat) that takes n_c_1_w_1_h_1 ... n_c_k_w_k_h_k and output n_(c_1+...+c_k)_w*h. This is useful for multi resolution feature (i.e. hypercolumn) and saves one memory allocation.
I'll refactor this if no one beats me to it.

[antinucleon]

I am out side and I will probably add bilinear upsampling in upsampling symbol. For multi input upsampling, can we use two symbol to do that? I

[piiswrong]

Using to symbols will result in doubling the memory used.

[antinucleon]

K, I only use this for GAN and double memoir doesn't hurt me seems. Also for bilinear seems we need to do similar thing. I will use symbol first to make GAN work.

[winstywang]

seems good to me.
@piiswrong For hypercolumn, it is better to implement bilinear upsamping or use the deconv directly.

[vchuravy]

We could also implement bilinear upsampling with the deconvolution layer, but we would need a way to prevent changes to the weights. See dmlc/MXNet.jl#34 for a first implementation on the Julia side.