Adding the image crop dataset implementation, with refactor of serialization logic to reduce complexity and redundancy

[wli51]

Overview

This PR adds the new image dataset class that returns user-specified crops dynamically extracted from full-resolution image files. The backend engine was introduced in #21.

Adds:

1. src/virtual_stain_flow/datasets/crop_dataset.py the new crop dataset

Refactors:

This PR also includes a small refactor to simplify backend serialization logic, addressing design issues introduced in #11.

Why the Refactor?

The dataset is based on:

• an immutable file-index DataFrame wrapped by DatasetManifest (validation + image access), and
• a FileState layer that handles caching and lazy loading.

In #11, reproducibility was implemented at the Dataset level by serializing the file index and manually reconstructing DatasetManifest → FileState. This tightly coupled Dataset with internal parameters of those classes. As DatasetManifest and FileState gained configuration options, Dataset had to know and serialize all of them, creating unnecessary complexity.

This PR moves serialization and deserialization logic into DatasetManifest and FileState themselves. The dataset now only manages its own parameters:

def to_config(self):
    return {
        "file_state": self.file_state.to_config(),
        "dataset_param1": ...,
    }

@classmethod
def from_config(cls, config):
    return cls(
        file_state=FileState.from_config(config["file_state"]),
        dataset_param1=config["dataset_param1"],
    )

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[wli51]

Thanks @d33bs. Addressed all your comments that I can. Updated the example and testing as well.