chore: use memory store in tests

[ilan-gold]

I try to resolve the problem "what if our zarr tests didn't take so long and we also didn't write SO much data during testing" by moving to memory stores where possible - there is still a test explicitly for local storage in tests/test_readwrite.py::test_readwrite_equivalent_h5ad_zarr (or where appropriate to not change) but otherwise, this should provide a big speedup!

• Closes #
• Tests added

• Release note not necessary because: testing change

[codecov]

Codecov Report

❌ Patch coverage is 96.15385% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 85.84%. Comparing base (d75540e) to head (d4f8a71).
✅ All tests successful. No failed tests found.

Files with missing lines	Patch %	Lines
src/anndata/tests/helpers.py	92.85%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2634      +/-   ##
==========================================
- Coverage   87.66%   85.84%   -1.82%     
==========================================
  Files          49       49              
  Lines        8121     8138      +17     
==========================================
- Hits         7119     6986     -133     
- Misses       1002     1152     +150     

Files with missing lines	Coverage Δ
src/anndata/experimental/merge.py	88.70% <100.00%> (+0.37%)	⬆️
src/anndata/tests/helpers.py	83.92% <92.85%> (-8.86%)	⬇️

... and 7 files with indirect coverage changes

[ilan-gold]

Not too shabby! https://github.com/scverse/anndata/actions/runs/33877901352/job/101039244161 for example took 13min but now it runs in 9min!

[flying-sheep]

very nice! I tried doing the same for HDF5, but it seems to give no advantage, so I just backported the helper from there.