fix: allow writing to 0-dimensional arrays with sharding

[NIK-TIGER-BILL]

Closes #3751

This PR fixes a bug where writing to a 0-dimensional array with sharding enabled would crash with an IndexError in the sharding codec's vectorized chunk slice lookup.

The root cause was that get_chunk_slices_vectorized in zarr/codecs/sharding.py assumed offsets_and_lengths was at least 2-dimensional, but for 0-dimensional arrays it has shape (2,) instead of (1, 2). This caused offsets_and_lengths[:, 0] to fail with "too many indices for array".

Fix: Add an early return in get_chunk_slices_vectorized to handle the 0-dimensional case by reshaping offsets_and_lengths to (1, 2) before slicing.

Test: Added test_sharding_zero_dimensional in tests/test_codecs/test_sharding.py as a regression test.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.29%. Comparing base (eac9c86) to head (bb315a8).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3966      +/-   ##
==========================================
+ Coverage   93.28%   93.29%   +0.01%     
==========================================
  Files          87       87              
  Lines       11745    11752       +7     
==========================================
+ Hits        10956    10964       +8     
+ Misses        789      788       -1     

Files with missing lines	Coverage Δ
src/zarr/codecs/sharding.py	89.32% <100.00%> (+0.23%)	⬆️

... 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.

[chuckwondo]

Thanks @NIK-TIGER-BILL!

Would you mind also adding a test that covers your patch that handles the 0D case? The test you added shows that the original case no longer errors, but that case no longer enters the function that you patched, so your patch itself remains untested, hence the failed codecov run.

[NIK-TIGER-BILL]

@chuckwondo Good catch — the integration test bypassed the patched function. I just pushed a direct unit test that exercises _ShardIndex.get_chunk_slices_vectorized with a 0-D offsets_and_lengths array, covering both the written and unwritten chunk cases. This should satisfy codecov for the new branch. Let me know if you would like anything else adjusted!

[chuckwondo]

Thanks @NIK-TIGER-BILL. Would you mind updating your branch and addressing the ruff pre-commit error? (See https://results.pre-commit.ci/run/github/48049137/1778641889.s58oPP4KQbGdn7pJdtUPAg)

@maxrjones and/or @d-v-b, please review.

[NIK-TIGER-BILL]

@chuckwondo Done — rebased on latest upstream/main and fixed the ruff unused-variable error (prefixed ends_e with an underscore). CI should be green now. Let me know if anything else is needed!

[NIK-TIGER-BILL]

@chuckwondo I believe I've addressed all feedback — rebased on latest upstream/main and fixed the ruff unused-variable issue (prefixed with ). Could you please take another look when you have a moment? Thanks!

[chuckwondo]

@d-v-b, this is a very small PR. Looks good to me, but would you mind reviewing as well?

[chuckwondo]

Thanks @NIK-TIGER-BILL!

[d-v-b]

this looks good! there's some baseline code smell in these classes but this fix is appropriately targeted. I'll open an issue with my follow-up recommendations.