Locality optimisations for multicore backend

[athas]

The multicore (and C) backends will sometimes generate inefficient code where arrays are sequentially traversed with a stride. This is obviously bad. I think a good solution would be to do something similar to the "kernel babysitter" used by the GPU pipelines, where we analyse traversal patterns and transpose the arrays in advance such that the eventual traversal will be optimal. This is not as good as tiling, but it is very general.