Fix neighbors transformer shortcut

[flying-sheep]

• Closes Leiden clustering returns different results for Scanpy 1.10.4 and 1.9.3 #3422
• Tests included or not required because:

• Release notes not necessary because:

I don’t get it, what does numpy mean here?

Not equal to tolerance rtol=1e-05, atol=0

Mismatched elements: 6 / 10000 (0.06%)
Max absolute difference: 5.96046448e-08
Max relative difference: 5.51590204e-14
 x: array([[0.      , 0.      , 0.      , ..., 0.      , 0.      , 0.      ],
       [0.      , 0.      , 0.      , ..., 0.      , 0.      , 0.      ],
       [0.      , 0.      , 0.      , ..., 3.39297 , 0.      , 0.      ],...
 y: array([[0.      , 0.      , 0.      , ..., 0.      , 0.      , 0.      ],
       [0.      , 0.      , 0.      , ..., 0.      , 0.      , 0.      ],
       [0.      , 0.      , 0.      , ..., 3.39297 , 0.      , 0.      ],...

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 75.50%. Comparing base (75c246d) to head (51a9e22).
Report is 3 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3444      +/-   ##
==========================================
+ Coverage   75.45%   75.50%   +0.05%     
==========================================
  Files         113      114       +1     
  Lines       13243    13282      +39     
==========================================
+ Hits         9992    10029      +37     
- Misses       3251     3253       +2     

Files with missing lines	Coverage Δ
src/scanpy/neighbors/__init__.py	80.31% <100.00%> (+0.10%)	⬆️
src/scanpy/neighbors/_backends/pairwise.py	100.00% <100.00%> (ø)
src/scanpy/neighbors/_types.py	100.00% <100.00%> (ø)

... and 1 file with indirect coverage changes

[scverse-benchmark]

Benchmark changes

Change	Before [75c246d]	After [51a9e22]	Ratio	Benchmark (Parameter)
+	18.6±0.5ms	24.4±0.9ms	1.31	preprocessing_counts.FastSuite.time_calculate_qc_metrics('bmmc', 'counts')
-	1.04±0.03ms	698±30μs	0.67	preprocessing_log.FastSuite.time_mean_var('pbmc3k', 'off-axis')

Comparison: https://github.com/scverse/scanpy/compare/75c246d191aaabb02e6cce9d926de9fb40dbc77e..51a9e22ae1f1d1f124858207777e55134c708962
Last changed: Thu, 23 Jan 2025 12:39:00 +0000

More details: https://github.com/scverse/scanpy/pull/3444/checks?check_run_id=36056403402

[flying-sheep]

Interesting, seems like pairwise is exactly as fast, but uses a lot more memory.

Since we only use it in the shortcut for small datasets, that’s not that big a deal, but one point against going back to pairwise

[flying-sheep]

this doesn‘t do anything when there is no duplicated data, so since we don’t support duplicated data, we don’t want this