Explanation pre- and post-processing pipeline

[aaarrti]

Description of the problem

• attribution preprocessing before metrics should be available as separate modules which could be pipelined, like torchvision.transform or sklearn Pipeline.
• At the moment, attribution preprocessing, such as normalise or interpolation, and even - attribution computation like softmax, are currently performed under-the-hood in each metric.
• is a follow-up for Question: attributions preprocessing and get_wrapped_model only accessible when computing attributions in metric #186