[ARM] Perform lane interleaving from reductions.
We have a pass for MVE to perform lane interleaving to make use of top/bottom
instructions, that adds shuffles before extends and after truncates. This
extends it to also start from add reductions, where the order of lanes does not
matter so the shuffle is not needed. We need to be careful about not breaking
the form of existing reductions, but otherwise can save some instructions and
awkward extends.
Differential Revision: https://reviews.llvm.org/D143396