(In)stability of rise imagenet tutorial output

[loostrum]

The heatmap produced in the rise imagenet notebook varies quite a bit when re-running the notebook due to the randomness of the masks. When someone runs the tutorial, they will probably compare to the existing output and expect it to be the same. We could fix this by either increasing n_masks, or by setting the numpy random seed.

A quick check shows that with n_masks = 10000 instead of 1000, the output seems stable. However, this makes the notebook run for several minutes which is not ideal either.