Downloading to S3 with TFDS errors out because gfile doesn't create virtual path on S3

Short description
The TFDS download manager uses the etils Tensorflow backend, but unlike in Google Cloud Storage which successfully creates virtual paths with no files within them, gfile.mkdir doesn't actually create a virtual directory on S3 unless there's an actual file to be written. This means that the check here fails even though gfile.makedirs(path) runs successfully. e.g.,

  File ".../lib/python3.8/site-packages/tensorflow_datasets/core/download/download_manager.py", line 39
8, in _download
    download_tmp_dir.mkdir()
  File ".../lib/python3.8/site-packages/etils/epath/gpath.py", line 180, in mkdir
    self._backend.mkdir(self._path_str, exist_ok=exist_ok)
  File ".../lib/python3.8/site-packages/etils/epath/backend.py", line 239, in mkdir
    raise FileExistsError(f'Cannot create dir. {path} is not a directory')
FileExistsError: Cannot create dir. s3://... is not a directory

Environment information

• Operating System: Ubuntu 20.04.5 LTS

• Python version: 3.8.10

• tensorflow-datasets/tfds-nightly version: tensorflow-datasets 4.8.2

• tensorflow/tf-nightly version: tensorflow-cpu==2.11.0

• Does the issue still exists with the last tfds-nightly package (pip install --upgrade tfds-nightly) ?

Reproduction instructions

You can verify that this is the issue by doing tf.io.gfile.mkdir(p) and then tf.io.gfile.isdir(p) for a new path on S3 and then a new path on GCS. On GCS isdir will return True, but on S3 it will return False.

Expected behavior
Unfortunately since S3 doesn't seem to create virtual paths when gfile.mkdir is called, checking with isdir will always fail; not sure if there's a nicer solution without handling the download to S3 differently or making the change in etils.

Start with tensorflow_datasets/core/download/download_manager.py at the _download call to mkdir, then inspect etils/epath/backend.py around the isdir check at line 238. Reproduce the difference with tf.io.gfile.mkdir and isdir on new S3 and GCS paths. Done means the S3 download path no longer fails on this directory check while preserving the existing GCS behavior.