Consistency improvements and tests

[gsarti]

Description

Changes taking place in this PR:

• A new method inseq.list_feature_attribution_methods() is introduced to provide a list of all names of available FA methods.

• Fixed the problematic setting in which layer attribution methods are combined with target-side attribution, now we raise an error to mark this setting as unsupported. This could be supported in the future.

• Removed the GradientShapAttribution method, since it leads to problematic behavior during tests, producing widely inconsistent results when evaluating attribution on a single input or a batch containing the same input (similar to DIG). The method can be reintroduced with an ad-hoc pull request addressing this issue.

• Huggingface models don't need the InterpretableEmbeddings helper class from Captum to perform attributions since they already support input embeddings for both the encoder and the decoder. In light of this, configure_interpretable_embeddings and remove_interpretable_embeddings were moved from the HuggingfaceModel class to the base AttributionModel class, with precisions on how to use them when extending the latter to new frameworks.

• Added a suite of test to verify consistent behavior for the model.attribute core method in the following settings:

  • Testing with and without GPU should produce the same results
  • Testing the same example inside a batch or as a stand-alone input should produce the same results.
  • Test that the attribution process completes successfully for all available methods and options (adding probabilities, target-side attribution, computing convergence deltas, force decoding with custom references)

Type of Change

• 🔧 Bug fix (non-breaking change which fixes an issue)