fix(hdf5): read AnnData >= 0.13 nullable-string-array columns

[balwierz]

Summary

CI on Ubuntu 24.04 went red (9 AnnData/HDF5 tests) because the runner picked up anndata 0.13, which changed the on-disk encoding of string obs/var columns and the DataFrame _index from a plain string-array dataset to a nullable-string-array group (a values string dataset + a boolean mask). vv's reader only knew the dataset form and silently skipped the group — so obs/var string columns, the row index (cell/gene identifiers), and the labelled X preview all came up empty on files written by current anndata.

vv now decodes the new form:

• read_anndata_dataframe's column walker and read_anndata_index_labels handle a nullable-string-array group via a shared read_nullable_string_array() helper, applying the mask as Arrow nulls.
• HDF5 bools decode to their enum member name ("TRUE"/"FALSE"), so the mask check accepts the string form as well as a genuine Arrow boolean.
• The legacy string-array dataset path is unchanged — both old and new files read.

Verification

• Reproduced by installing the latest anndata in a fresh venv (0.13.1) and regenerating the fixtures — 9 failures, matching CI. After the fix: full suite 374 against anndata 0.13 and 0.12 (backward + forward compatible).
• New tiny.nullstr.h5ad — hand-written with raw h5py, so the new path (including the mask → NA case, which the anndata-generated fixtures don't exercise: label='beta' is masked to null) is covered regardless of the installed anndata version.
• Committed anndata fixtures are left in the local form; CI regenerates them (now with 0.13) and the reader handles them.

This turns 24.04 green again (it was breaking every PR, not just this branch).

🤖 Generated with Claude Code