Allow for per-detector weight when making cell-based coadds

Final round of Jenkins: https://rubin-ci.slac.stanford.edu/blue/organizations/jenkins/stack-os-matrix/detail/stack-os-matrix/4310/pipeline

Some minor comments and one important question about whether/how we can assume the PhotoCalib is constant on the PRs.

In addition to making the behavior better, it turns out that computing the weight was a significant time sink in making the coadds. Using the mean variance from `finalVisitSummary` actually speeds up the task as well.

Waiting for Jenkins to be back up to run CI. In the meantime, `u/kannawad/DM-48649` collection in `/repo/main` has a `deepCellCoadd` for `tract=2234 AND patch=92 AND band='i' AND skymap='lsst_cells_v1'`

After some thinking, I think this would be best computed on calexps and stored in some form in the ccdVisitTable if not done already. This has the advantage of computing statistics prior to correlations introduced from warping, and maintains continuity across patch (and tract) boundaries.