Embeddings defaults to model.get_input_embeddings() for all models

[cdpierse]

• This is a small change but adds compatibility for most if not all models in the HF library when fetching embeddings
• This change is opinionated and now only selects input embeddings (word embeddings) as the embeddings from which it calculates attributions along.
• I'm still researching this a bit to sanity check this as a choice, there are subtle but real differences between the attributions when using input embeddings versus Bert like models model.bert.embeddings which is usually a collection of embedding layers.
• I'll convert this to a regular PR when I feel a bit more sure of the decision.

[cdpierse]

This feels ready to merge now. I've been reading how members of the Captum team are doing this and from what I've seen they are also commonly using word_embeddings/get_input_embeddings() for calculating attributions (see pytorch/captum#592 ).

There certainly are cases where embeddings w.r.t position and token type might be useful (layer conductance) but the library currently doesn't have support for that level of model introspection.