This PR is a little bit too big, but contains many of the last minute efforts before the demo presentation. There are three core components:

1. New Limb Masking - this was a rather difficult addition to enable the limb masking to occur at the model level, in the past we were doing this at the dataloader level but this was wasteful.
2. Nonlinear DL - nonlinear transforms were not attempted in past due to fear of scientific invalidation to my understanding. This option uses an invertable method to feature transform (simple log).
3. F10.7 - the plot to compare the value of the embeddings directly to something immediate. It received some push back in the meeting as to why would anyone want to model a proxy. The logic was to have a single slide/plot proving this embeddings capture enough information to recreate F10.7.