Interpretability bugs and improvements required

[jhnwu3]

Updating this is crucial to helping users understand how to define their own models to be compatible with the interpretability module.

• standardize the forward_from_embedding method in base model
• conditionally compute loss so the interpretability method don't need to worry about faking a label tensor
• masking has been proven to be problematic even for now. especially some methods will zero values for perturbation, which cause the x.sum(dim=...) == 0 method fail significantly.