Fix torch pad: (before, after) pairs were swapped for asymmetric pad_width

[Ryo-wtnb11]

torch_pad flattened pad_width and then reversed the whole flat tuple, which reverses each (before, after) pair as well as the axis order — so on the torch backend padding placed before an axis ended up after it, and vice versa. Symmetric pads are unaffected, which is presumably why it went unnoticed.

The fix is to reverse the axis order before flattening:

pad = tuple(itertools.chain.from_iterable(pad_width))[::-1]        # wrong
pad = tuple(itertools.chain.from_iterable(reversed(pad_width)))    # right

torch's flat format is (before_last, after_last, before_secondlast, ...), i.e. axes reversed but each pair in order. The single-tuple broadcast branch (len(pad) == 2 -> pad * ndim) also becomes correct for an asymmetric ((a, b),), which was previously swapped too.

test_pad only checked the output shape and sum, both of which are invariant under the swap. It now also compares against np.pad and has an asymmetric pad_width case; that test fails on the torch backend before this change and passes after.

Fixes #35

[jcmgray]

Looks good thanks!