Allow perturbation metric to have a separate function for evaluation of performance drop

[chengchihhong]

Currently, the perturbation loss metric assumes that the result is based on classification, so the performance drop is just the difference of output of softmax for a specific class, before and after perturbation.

To use it in other non-classification tasks, one needs to define what does it mean to be performance drop. This should be refactored as a separate function to be injected during the computation of the performance loss.

The default should be classification.

[chengchihhong]

In addition, it can be generalized not only to performance drop, but as an indicator on result deviation.