Extremely small weights in h5parmdb

[henedler]

I reported this issue before as part of #325, now I see it again - possibly unrelated to the other points in that issue.

When solving in mode 'rotation+diagonal', I get correct solutions, but the weights in the h5parmdb are extremely small:

        [[[5.0271739e-13, 5.0271739e-13],
          [5.0577533e-13, 5.0577533e-13],
          [5.0859501e-13, 5.0859501e-13],
          ...,

In my understanding, the h5parm weights are purely boolean flags of one or zero, yet here there are these very small floats.
I see this behavior in the up-to-date DP3v5 as well as in DP3v4.2. If I solve in mode 'diagonal' instead, all weights are perfectly fine. The behavior is the same no matter if I solve against MODEL_DATA or a sourcedb.

This issue seems to be only present under certain conditions. And example can be found here (~300MB):
https://cloud.hs.uni-hamburg.de/s/BPwA3eK859BanMZ

[rvweeren]

I also see this issue being present. This is in the master branch from a few days ago. The weights are around 1e-8 in my case. It seems that all rotation+diagonal solves suffer from this issue.