Predict multiple GSPs at once

[JackKelly]

This should help:

• increase the diversity of the training set
• improve national PV

Implementation

• in the data loader, pick a random gsp. Then find other GSPs within the ROI.
• pick a random subset of GSPs (so the example is centered randomly), up to some defined limit (say, 8).
• compute the centroid of all those chosen GSPs. That's the center of the example. (need to figure out how to "remember" the chosen GSPs)
• in the model, need to handle multiple GSPs, and mask missing ones.
• and need to give the GSP ID of each patch of satellite, and each PV system