Right now the way we canonicalize sparse arrays is super inconsistent (some manually build up arrays, some do a bunch of extra work and then patch, some use builders). This can be more consistent (and likely more efficient).