Add loading from the Datasets Hub + add relative paths in download manager

[lhoestq]

With the new Datasets Hub on huggingface.co it's now possible to have a dataset repo with your own script and data.
For example: https://huggingface.co/datasets/lhoestq/custom_squad/tree/main contains one script and two json files.

You can load it using

from datasets import load_dataset

d = load_dataset("lhoestq/custom_squad")

To be able to use the data files that live right next to the dataset script on the repo in the hub, I added relative paths support for the DownloadManager. For example in the repo mentioned above, there are two json files that can be downloaded via

_URLS = {
    "train": "train-v1.1.json",
    "dev": "dev-v1.1.json",
}
downloaded_files = dl_manager.download_and_extract(_URLS)

To make it work, I set the base_path of the DownloadManager to be the parent path of the dataset script (which comes from either a local path or a remote url).

I also had to add the auth header of the requests to huggingface.co for private datasets repos. The token is fetched from huggingface_hub.

[lhoestq]

I just added the steps to share a dataset on the datasets hub. It's highly inspired by the steps to share a model in the transformers doc.

Moreover once the new huggingface_hub is released we can update the version in the setup.py. We also need to update the command to create a dataset repo in the documentation

I added a few more tests with the "lhoestq/test" dataset I added on the hub and it works fine :)

[lhoestq]

Here is the PR adding support for datasets repos in huggingface_hub: huggingface/huggingface_hub#14