(sfnp) add permute host op.

[monorimet]

Some use cases require a bit of array handling in between shortfin program invocations or as pre/post-processing steps. In SD workflows, it's common to see a permutation of the image output array in preparation for image library (e.g. pillow)
consumption.

Perhaps implementing the equivalent of np.transpose (docs) would achieve what we need for most cases.

The following is a snippet showing how we use np.transpose in SDXL:

permuted = (
    np.transpose(vae_output, (0, 2, 3, 1))[0]
)