Convert: MTX and 10x MTX folders, import and export (streaming)

[Claptar]

Part of #20. Depends on #21. Equivalent of scanpy.read_10x_mtx / read_mtx.

Import

• 10x folders: matrix.mtx[.gz] + barcodes.tsv[.gz] + features.tsv[.gz] (or genes.tsv for Cell Ranger 2).
• STARsolo Solo.out/Gene/{filtered,raw} (same layout; features.tsv without the feature-type column).
• STARsolo velocyto output (Solo.out/Velocyto/raw/{spliced,unspliced,ambiguous}.mtx) → layers.
• A bare .mtx file, with optional --obs-names-file / --var-names-file.
• Orientation: 10x/STARsolo are genes × cells (transposed on read); --transpose / --no-transpose override.

Make it streaming. formats/sparse._read_mtx currently builds a Python list of every entry — unusable for large files. Replace with:

1. pass 1: stream the file (gzip as a stream), count nnz per output row → indptr
2. pass 2: scatter entries into preallocated on-disk indices/data, buffered in chunks
3. sort indices within rows only if the file isn't already ordered

The same reader should back adata import sparse, which benefits too.

Export

h5ad → 10x v3 folder with gzipped files: extend formats/sparse.export_mtx and write barcodes.tsv.gz / features.tsv.gz from obs / var. --layer picks the matrix.

Tests

• Fixtures for Cell Ranger 2, Cell Ranger 3, STARsolo (incl. velocyto) and a bare mtx.
• Compare with scanpy.read_10x_mtx / read_mtx (integration marker).
• Perf guard bounding reads/writes for the two-pass import.