[Experiment] Make S3Handler.s3_read return a stream rather than bytes

Originally I was expecting the returned stream from S3handler is non-seekable stream. But, it turns out that the whole archive/files will be dumped into memory based on the implementation (I might be wrong about it then I need someone to validate it)

• C++ side: https://github.com/pytorch/data/blob/a435c7f5f7543e3614130eb2ecee520396b8efd2/torchdata/csrc/pybind/pybind.cpp#L28
• Python side: https://github.com/pytorch/data/blob/a435c7f5f7543e3614130eb2ecee520396b8efd2/torchdata/datapipes/iter/load/s3io.py#L135
  And, that is the reason that the performance seems on parity with or without BytesIO in this issue.

In order to make it streaming, we need to have a way to pybind C++ stream IO to python, which is non-trivial. See a code example: https://github.com/CadQuery/OCP/blob/master/pystreambuf.h

Potentially this change would accelerate data preprocessing. But, it needs to be extensively benchmarked.

Start by reading torchdata/csrc/pybind/pybind.cpp around line 28 and torchdata/datapipes/iter/load/s3io.py around line 135 to verify whether S3Handler.s3_read materializes the archive in memory. Review the linked stream-binding example, then define benchmarks comparing current and streaming behavior. Done means a validated streaming approach with benchmark results showing its preprocessing impact.