Reduce memory footprint of `compute`

Currently, when `compute` is called all data is loaded into memory and passed as a list. This can pose a bottleneck especially for data intensive modalities (e.g. images) or measurements of large datasets. As an alternative we could pass a generator (or something similar) object that iterates over the datasets as many metrics are calculated in for loops or list comprehensions. If necessary we can still convert such an object to a list and load it into memory.

What do you think? @lhoestq @sashavor