feat(bids): Add BIDS dataset loader for neuroimaging data

[The-Obstacle-Is-The-Way]

Summary

Adds native BIDS (Brain Imaging Data Structure) dataset loading support using PyBIDS, enabling load_dataset('bids', data_dir='/path/to/bids') workflow for neuroimaging researchers.

Contributes to #7804 (Support scientific data formats) - BIDS is a widely-used standard for organizing neuroimaging data built on NIfTI files.

Changes

Core Implementation

• src/datasets/packaged_modules/bids/bids.py - GeneratorBasedBuilder implementation
• src/datasets/packaged_modules/bids/__init__.py - Module exports
• src/datasets/packaged_modules/__init__.py - Registration with module registry
• src/datasets/config.py - PYBIDS_AVAILABLE config flag
• setup.py - Optional pybids>=0.21.0 + nibabel dependency

Features

• Automatic BIDS structure validation
• Subject/session/datatype filtering via config
• JSON sidecar metadata extraction
• NIfTI file decoding via existing Nifti feature

Documentation & Tests

• docs/source/bids_dataset.mdx - User guide with examples
• tests/packaged_modules/test_bids.py - Unit tests (4 tests)

Usage

from datasets import load_dataset

# Load entire BIDS dataset
ds = load_dataset('bids', data_dir='/path/to/bids_dataset')

# Filter by subject/session
ds = load_dataset('bids', 
    data_dir='/path/to/bids_dataset',
    subjects=['01', '02'],
    sessions=['baseline']
)

# Access samples
sample = ds['train'][0]
print(sample['subject'])      # '01'
print(sample['nifti'].shape)  # (176, 256, 256)
print(sample['metadata'])     # JSON sidecar data

Test plan

• All 4 unit tests pass (pytest tests/packaged_modules/test_bids.py)
• make quality passes (ruff check)
• End-to-end tested with real OpenNeuro data (ds000102)

Context

This PR is part of the neuroimaging initiative discussed with @TobiasPitters. Follows the BIDS 1.10.1 specification and leverages the existing Nifti feature for NIfTI file handling.

Related PRs:

• Nifti visualization support #7874 (Nifti visualization support)
• Replace papaya with niivue #7878 (Replace papaya with niivue)

[CloseChoice]

I looked over the code and tested this and it looks absolutely fantastic. Also uploaded a dataset to test:

from datasets import load_dataset

ds = load_dataset("TobiasPitters/ds004884-mini")
ex = ds['train'][0]
ex['nifti']

or for streaming:

from datasets import load_dataset

ds = load_dataset("TobiasPitters/ds004884-mini", streaming=True)
ex = next(iter(ds['train']))
ex['nifti']

Here's how it's visualized:

@neurolabusc FYI

By the way using this branch (and niivue) I created: https://huggingface.co/spaces/TobiasPitters/bids-neuroimaging

[CloseChoice]

I think this might be something for another PR but actually having a dict-like object here would be more beneficial here. Not quite sure how we could achieve that, maybe through pyarrow's mapping and union type or having a dedicated feature for BIDSMetadata (or for dictionaries in general?).