Fix custom splits during preprocessing

[ChristianBaeuchl]

Summary

This PR fixes the handling of user-provided custom_splits during preprocessing.

Changes

_utils.py

• Added custom_splits_from_anno() function.
  • Reads an annotation file and converts a column containing split labels into the dict format expected by the splitter.
  • Sorts the annotation table by the sample_id column to align the resulting row indices with the sample IDs that are sorted during preprocessing.

_base_pipeline.py

• Updated BasePipeline.
  • Passes custom split information to the preprocessor used by the base pipeline.

_base_preprocessor.py

• Updated BasePreprocessor.
  • Added a custom_splits attribute to the constructor.
  • Passes the custom_splits information to PairedUnpairedSplitter.

_datasplitter.py

• Updated PairedUnpairedSplitter.
  • Added a custom_splits attribute to the constructor.
  • Updated split() so that it checks whether custom splits exist.
  • Added _per_modality_splits_from_custom_splits().
    • Checks custom split existence and required keys.
    • Takes the annotation index as the source of sample IDs.
    • Converts each split from row indices to sample-ID sets.
    • Checks for overlap between splits.
    • Creates empty split containers for every modality.
    • Maps custom split IDs to each data modality.
    • Checks whether any custom-split samples are missing from all modalities.
    • Returns sample-ID sets per modality.

Testing

Tested locally with:

• an Imagix IMG_TO_IMG example using 2D MedMNIST PNG data and an annotation file with a custom_splits column
• an Imagix3D IMG_TO_IMG example using predefined annotation-based train/validation/test splits

In both cases, preprocessing used the custom split path via PairedUnpairedSplitter.