Rename matrix convert to recast, and add the adata convert format framework #21

Part of #20.

Why

adata convert currently means "change a matrix's dtype, layout or density" (commands/convert.py, core/convert.py). Format conversion is the more natural meaning of the name, and users will look for it there.

Plan

1. Add adata recast, identical to today's convert. Move commands/convert.py → commands/recast.py and core/convert.py → core/recast.py; move tests (test_convert.py) and the COMMANDS.md section with them.
2. Transition period (0.7.x until 1.0, the same release that drops the h5ad alias):
   • adata convert called with the old matrix flags (--dtype, --layout, --dense, --sparse, --inplace) on AnnData → AnnData of the same backend forwards to recast and prints a deprecation warning.
   • Otherwise it runs format conversion.
   • Ambiguous case — backend change (h5ad → zarr) and matrix flags: convert the format, then recast, and warn.
3. Framework: the reader/writer registry, sniffing, the ConversionPlan dataclass and the common options described in the epic. Ship it together with the h5ad ↔ zarr converter (Convert: h5ad ↔ zarr #22) so the framework has one real user from day one.
4. Update CHANGELOG.md, docs/COMMANDS.md, README, and keep test_docs_are_accurate.py green.
5. Small fix on the way: _require_zarr still suggests uv sync --extra zarr, but that extra no longer exists.

Open questions

• Is recast the right name? Alternatives: retype, reformat, matrix.
• Should format-specific options live on convert directly, or be namespaced (--10x-genome, --loom-obs-names)? Flat options are friendlier; namespacing avoids clashes as formats grow.