Adding SEN12MS training task

[calebrob6]

Adds training support for a basic task involving the SEN12MS dataset. Train/val/test splits happen geographically, labels are the simplified MODIS IGBP labels (the first channel of labels provided in SEN12MS) used in the DFC2020 challenge.

To dos:

• Add segmentation-models-pytorch library to spack (https://github.com/qubvel/segmentation_models.pytorch)
• Figure out where to put trained model weights
• Add S2 only training as an option to the task

[adamjstewart]

Added tests, should be safe to merge now.

Hopefully it's obvious how to add tests for any new tasks you add (just add to the parametrize line). I had to modify my 2x2 test images to have the correct dimensions and band count in order for the tests to work.

I'll add a Spack package for segmentation-models-pytorch. We seem to be at the point where every new PR adds 1+ new dependencies, so we should be careful not to add too many deps.

[calebrob6]

I'll add a Spack package for segmentation-models-pytorch. We seem to be at the point where every new PR adds 1+ new dependencies, so we should be careful not to add too many deps.

Agreed, segmentation-models-pytorch seems to include timm as a sub-dependency so the only other dependency I see from the training side is some object detection model set.