Add VolumetricBatchNormalization

[colesbury]

All of the BatchNormalization modules now extend nn.BatchNormalization and use the same THNN/THCUNN implementation.

[fmassa]

I really like this PR, very clean and nicely done.
I have one question though, in the case one wants to use THNN as a standalone library: I think it's not intuitive that one should always feed a 3D tensor to BatchNormalization (even if we are in spatial or volumetric domain), and resize the output tensor afterwards so that it has the right size for the next module.
I was wondering if one couldn't change this part to compute n in a for loop over the dimensions of the input which are > 2. This solution might not apply as easily to the cunn version though, as the input dimensions are statically determined...

Maybe an easier solution (that could work for both backends) would be to do the resizing of the input/output inside the C/Cuda functions ?

[colesbury]

@fmassa I think that's a good idea. The tensor:view() function is implemented in Lua but not in C. I think we can effectively do the same thing in checktensor in the CUDA version.

[colesbury]

Updated PR with @fmassa's suggestion. (Tests pass VolumetricMaxPooling is flaky)

[fmassa]

Very cool, thanks @colesbury !

[hughperkins]

Hmmm, this breaks BatchNormalization in clnn :-( For example, this means that http://torch.ch/blog/2015/07/30/cifar.html will no longer run on clnn

[fmassa]

@hughperkins The previous commit (adding C/Cuda implementations of SpatialBatchNormalization) also is non-clnn friendly. Can't we (at least temporarily) monkey-patch clnn with the previous lua implementation ? (But we should pay attention because the meaning of some variables have changed since then)

[hughperkins]

Yes. "we" can :-P However, it would be nice to get a little @hughperkins in any future pull request that strips out lua implementation, and replaces it with c implementation :-)