Allow Fortran-style memory layout

[Markus-Goetz]

Feature functionality
Allow that tensors have Fortran-style, i.e. column-first, data layout. PyTorch supports this feature now directly via a passable parameter. This issue requires to work through factory functions, the creation of layout conversion calls as well as an investigation if and where auto-conversion between C- and Fortran-style layout needs to happen.

[ClaudiaComito]

@Markus-Goetz

Just checking if we are on the same page. For ndarrays, order='C' (default, rows first), or order='F' (columns first). So what we're talking about is to allow for a ht.tensor attribute order so that the memory layout of the tensor can be specified (or modified). E.g.:

a = ht.zeros((3,4,5), order='F')

would create a tensor with Fortran-style memory layout. We would then have a strides attribute to check the layout.

(Numpy details here)

I'm not sure about what PyTorch parameter you're referring to. With PyTorch 1.2.0 we can specify the memory layout (they call it memory_format) when creating a torch tensor, but this is with 4d tensors in NCHW order - <batch, channels, width, height> - in mind. It's not what we want.
pytorch/pytorch#19092

Then there's the torch.Tensor attribute layout, the only two options are torch.strided (default) or torch.spare_coo, so this isn't what we need either.
https://pytorch.org/docs/stable/tensor_attributes.html#torch-layout

So far, within torch the only way I've found of having column-major data layout is:

t = torch.randn(3, 5)
t = t.t().contiguous().t()

or

t.set_(t.storage(), t.storage_offset(), t.size(), tuple(reversed(t.stride())))

(from here)

ADDENDUM:
the torch option t = t.t().contiguous().t() only works with 2d tensors.