Add plot method and band selection to Zueri crop dataset

[nilsleh]

This PR adds the plot method as well as the ability to select bands when instantiating the dataset. The dataset is a timeseries dataset with 9 channels from Sentinel satellite imagery.

Changes:

• add band selection
• normalization
• plot method
• changed fake dataset to have 9 channels to test band selection

Issues/Questions:

• I was not able to find the specific band names the authors used so I contacted one of them and he stated that the band order is "NIR", "B03", "B02", "B04", "B05", "B06", "B07", "B11", "B12" and that the bands need to be normalized by 1e12, however, this yields tiny values and I read here that 1e4 is used to do the normalization
• the dataset contains mask, boxes, and labels. Which of these should be plotted?

Example with draw_segmentation_mask:

Alternative:

[isaaccorley]

• That's a strange band ordering. I would expect it to just be the original Sentinel-2 band ordering.
• You can divide Sentinel-2 by 1E4, however you may still find that it results in overly dark imagery. If so, just use our percentile normalization function.
• ZeuriCrop is a instance segmentation dataset so you can just plot image and masks. The masks should already be of shape (num_instances, h, w) so you should be able to use torchvision's draw_segmentation_masks directly.
• I extracted boxes from the instance masks only because torchvision's MaskRCNN expects boxes along with masks. However, you don't need to plot the boxes.