[Feature] Native Pytorch integrations

[jaychia]

Summary

Pytorch is used for distributed training, model inference etc. The most common abstraction used for data ingestion is a Pytorch Dataloader

Machine learning is an integral part of the Daft ecosystem, and having integrations with Pytorch will be extremely useful for Daft's users.

Read from Pytorch Dataset

Pytorch provides some datasets out-of-the-box:

1. https://pytorch.org/vision/stable/datasets.html
2. https://pytorch.org/text/stable/datasets.html
3. https://pytorch.org/audio/stable/datasets.html

Reading these datasets into a Daft dataframe, and providing an easy API for loading any Pytorch dataset into Daft, would be very useful for performing pre-processing using a DataFrame API

Daft Dataframes .to_pytorch_dataloader()

Daft Dataframes can provide a Pytorch Dataloader API for performing distributed training or inference. This should work well for both local and distributed training. This needs to be scoped out more!