Support obsm embeddings in the Loader

[moinfar]

Summary

Adds support for loading dense obsm arrays (e.g. a scVI embedding "X_emb") through the mini-batch Loader, yielded row-aligned with X.

• New obsm_keys argument on Loader.add_adata / add_adatas (and an obsm argument on add_dataset / add_datasets) selects which obsm keys to load — only the requested keys are read, keeping I/O minimal.
• Backward compatible: on-disk store format is untouched; existing calls are unchanged. The only observable difference is the extra obsm=None key in each yielded batch.
• Validation raises clear errors for: non-dense obsm, mismatched n_obs, and inconsistent keys / feature-shape / dtype across datasets. obsm dtype/shape is required to be homogeneous across datasets (kept simple; can be relaxed later).

Use of AI agents

This change was developed with AI assistance (Claude Code). I briefly reviewed all changes (the implementation and tests were authored via the tool and reviewed by me). The correctness of row alignment, concurrency, and edge-case handling was additionally checked by a second run of an AI agent.

[codecov]

Codecov Report

❌ Patch coverage is 94.02985% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.60%. Comparing base (6269ade) to head (dac0578).

Files with missing lines	Patch %	Lines
src/annbatch/loader.py	93.93%	4 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #252      +/-   ##
==========================================
- Coverage   93.29%   91.60%   -1.70%     
==========================================
  Files          15       15              
  Lines        1432     1489      +57     
==========================================
+ Hits         1336     1364      +28     
- Misses         96      125      +29     

Files with missing lines	Coverage Δ
src/annbatch/types.py	100.00% <100.00%> (ø)
src/annbatch/loader.py	88.53% <93.93%> (-2.40%)	⬇️

... and 3 files with indirect coverage changes

🚀 New features to boost your workflow:

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

[moinfar]

Is related to #91 and #102

[selmanozleyen]

Thanks! this will be done in the following PR's after #251, because we need to warn the users about the changing default behaviour first.

[moinfar]

Thanks a lot.

[ilan-gold]

Yeah, the goal is not to have to special-case this but to use https://anndata.scverse.org/en/stable/accessors.html

i.e.,

Loader(...).add_adatas(adatas, acc={A.obsm["pca"], A.layers["counts"]})

[moinfar]

This accessor concept is really interesting.
Looking forward to seeing it in the annbatch interface.