use AccumulatorMeanStack for coadds

[esheldon]

We can't do it now, but from Arun:

Reading in the images is required to bring them to the same zero point before coadding (lines 420-421). There's work happening just this week and in the coming weeks to make all those calibrations on the calexp after which those lines would not be required. We can switch to passing the refs to make_coadd then, but let's leave this in as is for this PR.

[arunkannawadi]

I don't think this is an issue, or at least, the title is very misleading. The task still uses AccumulatorMeanStack for coaddition. It is just that all of the images are currently being read before passing them to AccumulatorMeanStack (but only as subimages). I'd mark this as duplicate of #6 and close this one.