Add IterableDataset.decode with multithreading

[lhoestq]

Useful for dataset streaming for multimodal datasets, and especially for lerobot.
It speeds up streaming up to 20 times.

When decoding is enabled (default), media types are decoded:

• audio -> dict of "array" and "sampling_rate" and "path"
• image -> PIL.Image
• video -> torchvision.io.VideoReader

You can enable multithreading using num_threads. This is especially useful to speed up remote
data streaming. However it can be slower than num_threads=0 for local data on fast disks.

PS: Disabling decoding is useful if you want to iterate on the paths or bytes of the media files
without actually decoding their content.

Example: Speed up streaming with multithreading:

>>> import os
>>> from datasets import load_dataset
>>> from tqdm import tqdm
>>> ds = load_dataset("sshh12/planet-textures", split="train", streaming=True)
>>> num_threads = min(32, (os.cpu_count() or 1) + 4)
>>> ds = ds.decode(num_threads=num_threads)
>>> for _ in tqdm(ds):  # 20 times faster !
...     ...

why not multiprocessing ? decoding is done with the GIL released in soundfile/PIL/torchvision so multiprocessing would just use more memory

TODO

• test
• add to docs

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.