Fix dimension mismatch and add auto-loss support to TRAK

[traderbxy]

Description

Summary

This PR fixes a dimension mismatch issue and enhances the TRAK method's usability by adding automatic loss/probability function support for common task types.

Changes

1. Fix dimension conflict in sample_inputs

   • File: dattri/dattri/algorithm/block_projected_if/core/compressor.py
   • Issue: sample_inputs[0] had a dimension mismatch with model expectations, causing runtime errors.
   • Fix: Adjusted data preprocessing logic to ensure proper dimension alignment between sample_inputs and model input requirements.

2. Enhanced TRAK method with automatic function support

   • Modified classes: TRAKAttributor and AttributionTask
   • New parameter: Added task_type parameter supporting 'image_classification' and 'text_classification' values
   • Improvements:
     • When task_type is specified, users no longer need to manually design loss_func and correct_probability_func
     • Added built-in, optimized functions for both classification tasks
     • Extended support for dictionary-formatted inputs in text classification scenarios
     • Maintains backward compatibility - manual function definitions still work when task_type is not specified

Benefits

• Reduced complexity: Users can now use TRAK for common tasks without custom function implementation
• Better UX: More consistent API with other attribution methods
• Extended compatibility: Better support for text classification pipelines with dictionary inputs
• Bug fix: Resolves dimension-related crashes in the block_projected_if algorithm

[jiaqima]

Could you add a unit test related to this bug fix so that we could make sure future changes won't introduce such bugs again?

[jiaqima]

In addition, I would recommend separating this bug fix from the other change to maintain a clean change log.

[jiaqima]

Is F used?