Support _Conv.data_format for visualize methods

[Qinusty]

Currently visualize_cam and visualize_saliency assume models follow the standard keras.backend.image_data_format() which leaves the function unusable with models which aren't using this.

See saliency.py L174 for an example of usage here.

Potentially adding a data_format parameter or extracting this information from the penultimate_layer with penultimate_layer.data_format .