Our paper includes an I/O performance benchmark consisting of 114 Landsat 8 Level-2 scenes and 1 CDL 2019 mask. Since its publication, we have not been benchmarking I/O during PRs or between releases. This has led to difficulty in properly evaluating proposed changes (#1881) and massive performance bugs that have gone undetected for years (#1968).

This PR adds a dataset to properly test this. It consists of a single Landsat 9 scene and CDL mask with the following splits:

• original: the original files as downloaded from USGS Earth Explorer and USDA CropScape
• raw: the same files with compression and with CDL clipped to the bounds of the Landsat scene
• preprocessed: the same files with compression, reprojected to the same CRS, as COGs, with TAP

These roughly correspond to the same categories as used in our paper, with a few changes:

• Only a single Landsat scene
• CDL is clipped to the bounds of Landsat

These changes are made to make the dataset as small as possible so it can be quickly downloaded and fit on any system. I believe that the dataset is still useful even at such a small size. This should be considered to be version 1 of the dataset, with many future changes to evaluate a broader set of conditions.

• Add dataset
• Add data module
• Add trainer
• Add tests
• Add documentation

Sample image over Champaign County, IL, USA:

This PR also represents an interesting and novel dataset design. Note that instead of writing a custom RasterDataset and overriding __getitem__ to handle both Landsat and CDL files, I'm actually subclassing from IntersectionDataset, instantiating Landsat and CDL classes, and computing the intersection. This is significantly easier and could be used to simplify our implementations of other raster datasets with both images and masks: AgriFieldNet, Chesapeake, EnviroAtlas, GlobBiomass, L7 Irish, L8 Biome, LandCover.ai, etc.

FYI @yichiac

Closes #190