FRML-146 Migrate Preprocessing toward dataset for independent dataset processing

[Eve-ning]

Currently, the preprocessing step for our image and the label encoding are tightly coupled together, which causes large complicated workarounds. For example, leveraging the on_before_batch_transfer in our LightningModule inherited classes, then within that function, calling a 2-deep function preprocess, which made the complexity skyrocket.

We believe that the preprocessing should live within the dataset, specifically, the image processing.

However, we realized that if we preprocessed the tree labels in the dataset, results (as numbers) wouldn't make sense. Therefore, we're still opting to encode and decode that within the model, which makes sense in the long run as it can be then inferred standalone.

Major Changes

• Preprocessing step for the input images (x) are now done in the dataset. Encoding/decoding of labels still remain in the model
• The preprocessing step in on_before_batch_transfer is vastly simplified to just nan-mask out missing y labels
• Construction of the StandardScaler and OrdinalEncoder are now done internally in the FRDCDataset and FRDCModule respectively
• Introduced FRDCModule to reduce redundancy between MixMatch and FixMatch. This module handles:
  • Model Saving/Loading
  • X_step code signature consistency, due to our batch being an uncommon schematic
  • on_before_batch_transfer for nan-mask preprocessing of missing y labels
• Explicitly use the same StandardScaler from training to be transfered to validation and test sets within the training scripts