Check permutation map for convolution operations

[mdgrs]

When computing the (multi input / multi output) kernel matrix for a convolution operator, it is advantageous to permute the rows of the matrix to reduce the number of non-zero diagonals. This in turn reduces the number of diagonals that have to be extracted to compute the MatVec.

The h and w were interchanged. This is a problem if the filter is non square.

I'm also noticing an error in flattening the final output in the ISL string.

Originally posted by @asraa in #2941

[j2kun]

@mdgrs can you explain what situation uses a non-square filter? I don't recall seeing that in any examples.

[mdgrs]

I have not yet encountered a non square filter, so this was more of an observation. I opened a PR to make the change (and a test that fails without the fix)

#2986

For context, pytorch does support non square filters, so this is not purely academic: https://docs.pytorch.org/docs/2.11/generated/torch.nn.Conv2d.html