`sc.pp.scale` dask differs from csr sparse

[scverse-bot]

The daily CI failed on test for scanpy failed. Please go to the logs of the integration testing repo to review. @scverse/anndata

[ilan-gold]

This test and others similar ones fail now semi-regularly. I am genuinely not exactly sure what changed because it seems this happens with and without clipping/zero-center.

[ilan-gold]

Not sure about the milestone since this is coming from #3351 which is marked as 1.12 which I also dont fully understand

[ilan-gold]

New, worse failures, even after the tolerance was changed: https://github.com/scverse/scanpy/actions/runs/15209499566/job/42780110674?pr=3643