Permuting after loading from `PermutedDimsArray` for coalesced access

[AntonOresten]

I was messing around with a batched matmul kernel, and noticed a slowdown of up to 2-4x when passing PermutedDimsArray. I then made a version that took the underlying array and permuted the tile after loading, which was way closer to the contiguous baseline. I assume the difference is that it was able to load with coalesced access and permute within registers.

After #69, would it be possible to use type information like StrideDivBy to load non-contiguous data as contiguous and permute after the load itself? I'm not sure how this relates to a sentence from the docstring: stride_div_by[i] = 4 means stride[i] is divisible by 4 (enables vectorized access), and what optimizations can already be made by the compiler.

[maleadt]

No, StrideDivBy is only there to indicate that contiguous memory can be loaded in a vectorized way, which PermudeTimsArray much certainly is not. That said, apparently Tile IR has an order argument to load and store indicating how logical tile dimensions map to physical memory dimensions. I think we could use that, but it's currently not exposed.

[maleadt]

Actually, order can be used to lazily do permuted loads, but it would probably be too fancy to forward that from PermutedDimsArrays inputs over ArraySpec to load/store automatically. Here, I think there's not much to be done; passing non-contiguous memory is going to be slow, and it's up to the programmer to instead pass the underlying contiguous memory and perform the permutation lazily, I guess.