Contant time len() will probably require additional storage. Atleast for my usecase that would be quite unfortunate as my program has hundreds of thousands of bitmaps at any time.

Perhaps instead the name of len() could be changed to cardinality() or another name so it doesn't conflict with the API guidelines and doesn't bloat RoaringBitmap with extra bytes?

#127 removes 8 bytes from each array container, so maybe we can afford a few extra bytes at the root now ;)

This is relevant to #167.

Cardinality of all ops can all be trivially implemented in terms of the cardinality of the intersection |A ∩ B|.

|A \ B| = |A| − |A ∩ B|
|A ∪ B| = |A| + |B| − |A ∩ B|
|A △ B| = |A| + |B| − 2|A ∩ B|

Intersection cardinality is cheaper to compute, so it's both simpler to implement and has better runtime performance if we have the len cached.