Update EuroSat eval task to use splits from https://arxiv.org/pdf/1911.06721.pdf

[gabrieltseng]

This PR updates the EuroSat splits so that they reflect the splits used by In-domain representation learning for remote sensing.

These are the splits also used by TorchGeo, allowing for easier comparison to TorchGeo models.

The updated accuracy results (Table 5 from the paper) are below. The table compares the original splits (what is currently in the paper) with the new splits (in this PR):

Model	# input pixels	k=5	k=20	k=100
ScaleMAE @ 16	256	0.729	0.727	0.695
SatMAE @ 16	256	0.723	0.721	0.676
Presto (RGB), original splits	9	0.824	0.804	0.752
Presto (MS), original splits	9	0.864	0.841	0.795
Presto (RGB), new splits	9	0.815	0.800	0.760
Presto (MS), new splits	9	0.860	0.840	0.794

[gabrieltseng]

Superseded by #14