Directly reading parquet files in a s3 bucket from the load_dataset method

### Feature request

Right now, we have to read the get the parquet file to the local storage. So having ability to read given the bucket directly address would be benificial

### Motivation

In a production set up, this feature can help us a lot. So we do not need move training datafiles in between storage.

### Your contribution

I am willing to help if there's anyway.