Generate masks on-the-fly in batches

[loostrum]

When using many masks, pre-generating them all could use too much memory.
In the dianna-deeprank branch (see

dianna/dianna/methods/rise.py

Line 190 in c1823d7

self.avg_mask = np.zeros(img_shape)

), we implemented a way to generate masks on the fly for RISE. This may be useful to add to the general DIANNA code.

It works as follows:
Masks are generated per batch. A running average is used to update the RISE output after running it with a new batch of masks.
Additionally we implemented an early stopping feature. The number of masks is no longer a set number that is used, but rather a maximum. When the output of RISE doesn't change anymore (to some tolerance), it stops and reports how many masks were actually used.