[FEAT] dataset from directory improvments

There are a few improvements that can be added that would facilitate the use of this functionality.
No need to do all of them at the same time

1. Add an additional function of image_dataset_from_directory that create vision data with task type "other" that can be really useful for image integrity and train test validation for all tasks. Can work on the following directory structure:
- root/
- train/
image1.jpeg
- test/
image1.jpeg

2. Change the function API to use generic generator instead of torch thereby removing the dependency in torch.
3. Support additional image formats

Start by locating the existing image_dataset_from_directory implementation and its tests. Review the requested directory structure and decide whether to scope work to the "other" task type, replacing the torch-based generator, or adding image formats; done should be defined by the selected improvement and corresponding coverage for that behavior.

Written by the indexing model from the issue text.