Add auto-shard configuration and manage anndata/dask pins

[lorenzocerrone]

Fix

• Replace da.to_zarr with da.store(..., lock=False) in pyramid writes (_on_disk_dask_zoom, _on_disk_coarsen) and region slice writes (_ops_slices). Dask >=2025.11's to_zarr re-derives chunks via normalize_chunks(chunks="auto", ...) and emits a PerformanceWarning (treated as error by ngio's filterwarnings) when the result is not a multiple of the target's chunks; da.store writes blocks 1:1.
• Copy object/string-dtype zarr arrays directly when consolidating groups: dask >=2025.11 raises NotImplementedError from auto-chunking for these dtypes, so they bypass dask and are copied via numpy.
• Set auto_shard_zarr_v3 together with zarr_write_format on anndata's global settings via a new _update_anndata_global_settings helper, so reading/writing tables works correctly when mixing zarr v2 and v3 in the same session on anndata 0.12.

Chores

• Pin anndata to >=0.12.0,<0.13.0.
• Unpin dask (remove the <2025.11.0 upper bound introduced in v0.4.5).

Checklist before merging

• I added an appropriate entry to CHANGELOG.md

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!