Add support to return dataset as a HF Dataset object in load_dataset()

[Kostis-S-Z]

load_dataset() can now return datasets, also as HuggingFace datasets objects using a new return_format argument. pandas DataFrame still remains the default, so this is an optional feature added on top.

from datacollective import load_dataset                                                                                                                                                                                                                                                                           
                                                                                                                                                                                                                                                                                                                    
ds = load_dataset("your-dataset-id", return_format="hf")                                                                                                                                                                                                                                                          

Changes

• New return_format="pandas" | "hf" argument on load_dataset (default "pandas", fully backward compatible). With "hf", single-split datasets return a Dataset; multi-split datasets (e.g. Common Voice) return a DatasetDict keyed by split name, driven by the dataset schema's splits field.
• Optional dependency: the HF datasets library is relatively heavy, so in order to keep our client lightweight, it will ship as an extra, i.e. uv add / pip install "datacollective[hf]". Without it, return_format="hf" raises a MissingDependencyError with install instructions before any API call or download, so users never hit the error after pulling a large archive.
• New hf_utils.py module: loaders are untouched and still produce DataFrames. The conversion happens post-hoc, so both formats always contain identical data. datasets is imported lazily and only referenced in type annotations under TYPE_CHECKING.
• Docs: new "Return as HuggingFace Dataset object" section in index.md. Also added a section on how to lazily decode audio with cast_column("audio", Audio(...)) which might be useful for all the ASR / TTS datasets! Plus, a README Quick Start step, and an HF example in demo_download_and_load.py that runs only when datasets is installed.
  • Tests: 12 new tests covering the conversion DataFrame -> Dataset (single split, multi-split, category dtypes), the missing-dependency error and fail-fast ordering, argument validation, and the end-to-end load_dataset for both formats.