Fix the implementation of ScoreCAM methods

[frgfm]

There are a few subtle differences between paper and the implementations in this repo. Some need to be fixed as they do not reflect any voluntary improvement upon the author's original idea:

• in ScoreCAM, SS-CAM & IS-CAM, the softmax needs to be applied along the channel dimension at the final step https://github.com/frgfm/torch-cam/blob/master/torchcam/methods/activation.py#L170
• in SS-CAM & IS-CAM, there is a mistake as, by inheritance (https://github.com/frgfm/torch-cam/blob/master/torchcam/methods/activation.py#L191), inputs are already multiplied by the upsampled activations, and are again multiplied in each implementation