Convert: CellBender `remove-background` output h5

[Claptar]

Part of #20. Depends on the 10x h5 reader in #23.

Format

output.h5 / output_filtered.h5 are 10x-v3 compatible under /matrix, plus:

• /droplet_latents: barcode_indices_for_latents, cell_probability, cell_size, background_fraction, droplet_efficiency, gene_expression_encoding
• /global_latents: ambient_expression, empty_droplet_size_lognormal_loc, posterior_probs, …
• /metadata: learning curve, target fpr, and more

Mapping

• X ← denoised counts (via Convert: 10x HDF5 (Cell Ranger *_feature_bc_matrix.h5), import and export #23)
• per-droplet latents → obs columns, scattered by barcode_indices_for_latents; NaN for droplets not analysed
• gene_expression_encoding → obsm["X_cellbender"]
• ambient_expression → var["ambient_expression"]
• scalar globals and metadata → uns["cellbender"]

Options

• --raw RAW.h5 adds the raw Cell Ranger counts as layers["raw"], aligned by barcode (like cellbender.remove_background.downstream.load_anndata_from_input_and_output)
• --filtered keeps only barcodes with cell_probability > 0.5 when converting the unfiltered file (threshold configurable)

Gotchas

CellBender 0.3.0 wrote h5 files that scanpy failed to read until 0.3.2 added a ptrepack fix. Accept both 0.2.x and 0.3.x layouts.

Tests

h5py fixtures mimicking 0.2 and 0.3 outputs; verify against a real CellBender run from cellgeni tickets; compare X with scanpy.read_10x_h5 on the same file.