`InstanceNorm1d` does not match documentation and is redundant.

[veugene]

InstanceNorm1d expects only 3D inputs, while the documentation states:

"Applies Instance Normalization over a 2D or 3D input (a mini-batch of 1D inputs with optional additional channel dimension)"

Furthermore, InstanceNorm1d, InstanceNorm2d, and InstanceNorm3d appear to be redundant as they add nothing to their parent class _InstanceNorm except an input dimension check.

Proposal A:
Fix the dimension check in InstanceNorm1d to allow both 2D and 3D inputs.

Proposal B:
Merge all instance norm classes into a single class, for any input dimensions > 1.

[veugene]

BatchNorm1d, BatchNorm2d, and BatchNorm3d are similarly redundant.

[arogozhnikov]

(confirm that in 0.4 InstanceNorm1d doesn't work with 2D inputs)

[soumith]

let's go with proposal A. We planned to deprecate (not delete) InstanceNorm*d in favor of a universal InstanceNorm, but haven't gotten around to it.