add tests for reading of legacy files (closes #92)

[ilia-kats]

Reference files are generated with src/mudata/tests/reference.py. This is based on the mudata fixture in tests/conftest.py, but it had to be modified, since NumPy does not guarantee stability of its random number generator across versions, so this needs to be completely deterministic.

See also the discussion in #92 .

[codecov]

Codecov Report

❌ Patch coverage is 87.50000% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.11%. Comparing base (2147f72) to head (69606bf).

Files with missing lines	Patch %	Lines
src/mudata/tests/reference.py	87.09%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #168      +/-   ##
==========================================
- Coverage   91.12%   91.11%   -0.02%     
==========================================
  Files          11       12       +1     
  Lines        1826     1857      +31     
==========================================
+ Hits         1664     1692      +28     
- Misses        162      165       +3     

Files with missing lines	Coverage Δ
src/mudata/_core/utils.py	93.97% <100.00%> (-1.21%)	⬇️
src/mudata/tests/reference.py	87.09% <87.09%> (ø)

... and 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ilan-gold]

Main question is: why not use h5diff for the newer file versions? #92 (comment) seems to only say some subset of the older verisons would not work under this (I think only the v0.1.2?)

[ilia-kats]

v0.2.4 was generated with AnnData 0.10.8, so that would also be affected (although I guess I could regenerate with AnnData 0.11, but I specifically tried to use the oldest supported versions). Also from a conceptual point of view, we only care about reading old files, nobody guarantees that writing will generate the same file as some previous version, so h5diff doesn't seem like the right approach. Finally, I think this approach is more future-proof: Suppose MuData or AnnData add additional attributes in the future, say another mapping (for example, if we decide to integrate #111 into MuData). Even an empty mapping will create an empty group in the HDF5, making h5diff fail.

[ilan-gold]

Ok! Looks good then!