Add checkpointing support

[bernardhan33]

Add checkpointing support for the Dataflux Dataset.

From investigation it seems that we can confidently use blob.open() to create a file handler.

• blob.open("rb") will return a BlobReader.
• blob.open("wb", ignore_flush=True) will return a BlobWriter.

Both types implement io.BufferedIOBase so can be passed into torch.save() and torch.load() like demonstrated in the README.

TODO: I haven't added unit tests to this yet -- I thought about writing mocks but decided to go with fakes. I wanted to share this PR first before making changes to fake_gcs so it would make more sense there. Will follow up on this next.

• Tests pass -- tested by running the demo code directly and verified it works.
• Appropriate changes to documentation are included in the PR -- updated the README.