Prefetching for IterableDataset

### Feature request

Add support for prefetching the next n batches through iterabledataset to reduce batch loading bottleneck in training loop.

### Motivation

The primary motivation behind this is to use hardware accelerators alongside a streaming dataset. This is required when you are in a low ram or low disk space setting as well as quick iteration where you're iterating though different accelerator environments (e.x changing ec2 instances quickly to figure out batch/sec for a particular architecture).

Currently, using the IterableDataset results in accelerators becoming basically useless due to the massive bottleneck induced by the dataset lazy loading/transform/mapping.

I've considered two alternatives:
PyTorch dataloader that handles this. However, I'm using jax, and I believe this is a piece of functionality that should live in the stream class.

Replicating the "num_workers" part of the PyTorch DataLoader to eagerly load batches and apply the transform so Arrow caching will automatically cache results and make them accessible.

### Your contribution

I may or may not have time to do this. Currently, I've written the basic multiprocessor approach to handle the eager DataLoader for my own use case with code that's not integrated to datasets. I'd definitely see this as being the default over the regular Dataset for most people given that they wouldn't have to wait on the datasets while also not worrying about performance.