Support for TFRecords preparation in ImageFolder

Is your feature request related to a problem? Please describe.
tfds.ImageFolder replaces the legacy dataset named 'image_label_folder' in favor of a clean API that can be used to easily load datasets from local (or even GCS bucket) directories.

The download_and_prepare method:
https://github.com/tensorflow/datasets/blob/adb320ff04b6e93c561dacb2b647c8fcbfea92f3/tensorflow_datasets/core/folder_dataset/image_folder.py#L106-L112
present in ImageFolder doesn't provide the (undocumented) functionality of TFRecords preparation from custom directories unlike the deprecated 'image_label_folder', which used to well support that! It indeed encourages the users to directly load pipelines from the same directory structure without any overhead (very useful for smaller datasets) but the same might not be the desired option for medium-large datasets.

Describe the solution you'd like
Is there a clean (public) API available in TFDS for such easy TFRecords preparation from custom directories? I'd assume that interoperability of that with ImageFolder would greatly benefit users.

Describe alternatives you've considered
Are there any plans to support such a feature in the future? (if possible, we could implement the download_and_prepare method for ImageFolder referring to the implementation in older 'image_label_folder')

Start with tensorflow_datasets/core/folder_dataset/image_folder.py around download_and_prepare at lines 106-112, then compare the deprecated image_label_folder implementation for its TFRecords preparation behavior. Determine the intended public API and add coverage for preparing TFRecords from custom directories once the scope is agreed.