perf: speed up backed sparse integer indexing

[tippered1-debug]

Summary

Integer indexing of backed sparse datasets currently reads every selected row or column separately, even when most of the indices form contiguous runs.

This PR groups those runs and reads them as slices. Shuffled indices and duplicates are restored to their original order after reading. Highly fragmented indexers continue to use the existing path, since slicing does not help in that case.

Related to #1224.

Benchmark

Benchmarked on a 10000 x 10000 sparse matrix with density 0.01. Each indexer contains 2048 elements. Results are medians of three runs.

HDF5 CSR

• Single run: 7.10 ms → 0.73 ms — 9.7x faster
• Multiple runs: 6.81 ms → 0.74 ms — 9.2x faster
• Clustered, shuffled: 7.04 ms → 1.03 ms — 6.8x faster
• Clustered with duplicates: 7.02 ms → 0.83 ms — 8.5x faster

Zarr CSR

• Single run: 16.60 ms → 3.70 ms — 4.5x faster
• Multiple runs: 16.27 ms → 13.92 ms — 1.2x faster
• Clustered, shuffled: 18.91 ms → 14.30 ms — 1.3x faster
• Clustered with duplicates: 19.19 ms → 8.53 ms — 2.2x faster

Fragmented indexers continue to use the existing path and stayed close to the previous timings.

• Closes #
• Tests added

• Release note not necessary because:

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 85.71%. Comparing base (6d70d77) to head (653f5b6).
⚠️ Report is 8 commits behind head on main.
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #2506      +/-   ##
==========================================
- Coverage   87.61%   85.71%   -1.91%     
==========================================
  Files          49       49              
  Lines        7693     7719      +26     
==========================================
- Hits         6740     6616     -124     
- Misses        953     1103     +150     

Files with missing lines	Coverage Δ
src/anndata/_core/sparse_dataset.py	92.15% <100.00%> (+0.01%)	⬆️

... and 7 files with indirect coverage changes

[ilan-gold]

Can this reuse subset_by_major_axis_mask? It looks eerily familiar. Or could that method be removed in favor of this code path? I'm not sure what the performance considerations here are but I think they are probably trivial i.e., operations on a small 1d array.

Thanks for the contribution!