Download state data on demand in the ChesapeakeCVPR dataset

[calebrob6]

Currently, downloading the entire 140GB dataset zip file is prone to error out due to http errors. Potentially better solutions would be:

• Split up the zip file and have the Dataset object only download the files it needs (i.e. if a user only wants to run Delaware experiments, they will only need to download ~10-20GB)
• Use some azure library to directly download the files (148GB total)

Regardless, we should link to the azcopy download instructions (bottom of https://lila.science/datasets/chesapeakelandcover) in the docs so that users can easily try other download options if the download=True method isn't working.

@hannah-rae for visibility!

[nilsleh]

I would advocate for the first solution, since I have run into that issue as well, where a remote machine with GPU has not enough memory for all 140GB but the current code implementation always checks for all state files to be present. I can open a PR to change the code.

[calebrob6]

Okay I'll see about creating zip files for each state.

[adamjstewart]

Yes, this is also how the non-CVPR version of the dataset is distributed.

[adamjstewart]

Is this and #484 still a WIP or should we close them?