Mask flexibility for methods than need to sample in the masked region

[ethanweber]

I'm making minor changes to make the masking feature more flexible. I have a method where I want to sample everywhere in the image, including the masked out regions, as specified in the transforms.json file. My custom model will compute losses differently depending on where it is masked out.

Furthermore, I sometimes want to give the masked out region a specific color. A minor modification to the dataloading and dataset enables this.

Lastly, I expose the trainer to the callback because my pipeline needs access to this, specifically to read off the base_dir so I can avoid wandb.

[kerrj]

sgtm