feat: parallelize per-library spatial graph construction

[Marius1311]

Add n_jobs parameter to spatial_neighbors() to compute per-library graphs in parallel via joblib. Defaults to 1 (sequential, no behavior change). Set to -1 to use all CPUs.

When library_key is set, each library's graph is already computed independently, so this is a trivially parallel workload. For datasets with many libraries (e.g., multi-sample spatial transcriptomics), this gives a near-linear speedup.

IMPORTANT: Please search among the Pull requests before creating one.

Description

How has this been tested?

Closes

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74.08%. Comparing base (9690a55) to head (5eda868).
⚠️ Report is 6 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1144      +/-   ##
==========================================
+ Coverage   74.05%   74.08%   +0.03%     
==========================================
  Files          39       39              
  Lines        6495     6503       +8     
  Branches     1122     1122              
==========================================
+ Hits         4810     4818       +8     
  Misses       1230     1230              
  Partials      455      455              

Files with missing lines	Coverage Δ
src/squidpy/gr/_build.py	88.88% <100.00%> (+0.36%)	⬆️

🚀 New features to boost your workflow:

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

[selmanozleyen]

Thank you but we are working on making updates to spatial_neighbours: #1147 Stay tuned, after these changes your version should be trivial to implement!

[grst]

Hi @Marius1311,

#1147 has been merged.
Your implementation would have to be adapted accordingly, to parallelize this section:

squidpy/src/squidpy/gr/_build.py

Lines 793 to 799 in da789d0

	if library_key is not None:
	mats: list[tuple[Any, Any]] = []
	ixs: list[int] = []
	for lib in libs:
	ixs.extend(np.where(adata.obs[library_key] == lib)[0])
	mats.append(builder.build(adata[adata.obs[library_key] == lib].obsm[spatial_key]))
	adj, dst = builder.combine(mats, ixs)

Do you want to work on this?

[Marius1311]

thanks @grst! In my own tests, my parallelization actually didn't lead to any speed improvements, not sure why!

[Marius1311]

So we can close this if you like.

[grst]

I observed something similar, and I think a large part of the runtime is spent on subsetting the spatialdata/anndata object before any nearest neighbor code is even triggered.

I still think it should be possible to optimize this, but then it's likely not a trivial fix as thought initially.

[grst]

Closing in favor of #1198