feat: python client #45

[jdpearce4]

Summary

Introduce a first-class Python client (TranscriptFormerClient) for inference and artifact/data downloads. The client mirrors the CLI configuration behavior while providing a simple, programmatic API that returns an in-memory AnnData object.

Key Features

• In-memory inference: inference(...) returns an anndata.AnnData without writing to disk.
• Config parity with CLI: Builds a Hydra-compatible config by:
• Loading the same CLI YAML defaults.
• Applying dataclass overrides from kwargs (InferenceConfig, DataConfig).
• Merging with the checkpoint config.json via the same utility used by the CLI.

Convenience downloads:

• download_model(...) for checkpoints and embeddings.
• download_data(...) for CellxGene datasets by species.
• download_dataset(...) for curated sources (e.g., Tabula Sapiens, Bgee).
• Logging control: Optional log_level argument to run quietly or verbosely without affecting global logging.

API Surface

• TranscriptFormerClient.inference(data_file, checkpoint_path, **kwargs) -> anndata.AnnData
• Accepts most InferenceConfig and DataConfig fields as kwargs (e.g., batch_size, output_keys, gene_col_name, use_raw, use_oom_dataloader, n_data_workers, etc.).
• Returns a single AnnData with obsm/uns populated per output_keys.
• TranscriptFormerClient.download_model(model, checkpoint_dir=...) -> None
• TranscriptFormerClient.download_data(species=[...], output_dir=..., ...) -> int
• TranscriptFormerClient.download_dataset(dataset, ...) -> anndata.AnnData | None