Concurrent DCP S3 range read

[clumsy]

Tell us more about this new feature.

There's opportunity to significantly speed up Distributed Checkpoint (DCP) S3 reads using concurrent S3 range fetches. The idea is to load data the exact S3 range and then load the tensor/bytes into PyTorch in a dedicated thread using checkpoint metadata. This is especially beneficial for instances with large network bandwidth and big number of CPUs which allows to saturate I/O.