[NEON] Fix scalar path of saturating_absdiff()

In the previous version 'decltype(src_a - src_b)'
was still signed, therefore it chose the first
argument even when max() should have been chosen.

The new variant calculates unsigned difference and
then applies saturating cast to limit the domain.

Co-authored-by: Denes Tarjan <denes.tarjan@arm.com>