feat: wire scfit dataloader

[selmanozleyen]

This pull request refactors and simplifies the data handling utilities in the core methods, improves the interface for slicing and matching batch data, and updates dependencies and API usage for improved compatibility and clarity. The most important changes are summarized below.

Data handling and API refactoring

• Major simplification of src/sckitflow/core/_data_utils.py: removes several now-redundant functions (extract_step_data, extract_coupling_data, extract_distribution_data, etc.), and introduces a new, general-purpose subscript_step_data function for row-slicing both source and target fields in a StepData batch. This centralizes and clarifies batch manipulation logic. [1] [2] [3] [4]
• Refactors all internal usage to rely on subscript_step_data, removing the previous _safe_subscript_obj and custom slicing logic from BaseMethod. This makes the codebase more maintainable and less error-prone. [1] [2] [3]

Documentation and interface improvements

• Updates docstrings and comments throughout the codebase to reflect the new data loader responsibilities and clarify that StepData is now fully assembled by the data loaders, not by methods themselves. [1] [2] [3]

Dependency and configuration updates

• Updates minimum versions of key dependencies in pyproject.toml (anndata>=0.13, scfit>=0.0.7) to support new features and GPU streaming, and adds optional GPU and torch dependencies for improved flexibility. [1] [2]

Notebooks and example usage

• Updates example notebooks to use the new APIs: replaces optim_kwargs with optim_config, and updates batch size and prediction parameterization for consistency with the refactored interfaces. [1] [2] [3] [4]

Method registry and exports

• Exposes OptimConfig and OptimizationManager in the core methods module for easier access and consistency with the rest of the API.

These changes collectively improve the clarity, maintainability, and extensibility of the codebase, especially around how batches of data are handled and manipulated during training and inference.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB