Convert: loom (generic and velocyto), import and export

[Claptar]

Part of #20. Depends on #21. Equivalent of scanpy.read_loom / anndata.read_loom.

Import

loom	AnnData
/matrix (genes × cells, dense, chunked)	X (transposed)
/layers/*	layers
/col_attrs/*	obs (1-D) / obsm (2-D)
/row_attrs/*	var (1-D) / varm (2-D)
/col_graphs/*, /row_graphs/* (COO a, b, w)	obsp / varp
global attributes (/attrs in spec v3, root attrs in v2)	uns

Options

• --obs-names / --var-names choose the attribute (defaults CellID / Gene, as loompy)
• --sparse (default, like scanpy's sparse=True) / --dense
• --X-layer NAME promotes a layer to X
• --make-unique

Velocyto looms: spliced, unspliced, ambiguous (and matrix) → layers, ready for scVelo.

Streaming: read /matrix in column blocks aligned to its chunking (usually 64×64), sparsify each block and emit CSR rows; reuse sparsify / transpose_sparse_streaming.

Export

h5ad → loom for SCope, velocyto and loompy users.

• Dense transpose written in chunks; refuse above a size threshold unless --force (same idea as check_growth).
• obs/var → col_attrs/row_attrs; obsm/varm 2-D arrays kept; categoricals written as strings.
• obsp/varp → graphs.

Tests

• h5py fixtures for spec v2 and v3, and a velocyto-style loom.
• Compare with scanpy.read_loom / anndata.read_loom; round trip h5ad → loom → h5ad.