Specifically, previously, each MUC would be added as a roster item, and
then each join presence would be counted as a resource of that item,
triggering 1 to 5 events and more backend logic in slixmpp.
As a result, joining big rooms is tremendously slow, (JID() calls,
event() calls, __getitem__ calls for nothing), and takes RAM (a quick
tracemalloc tells me around 1 MiB for 3500 participants, i.e. 2 big IRC
rooms). Those resources may not necessarily be cleaned properly, leading
to memory leaks on long-term usage.
This is a micro-optimization that adds an attribute to roster items so
that MUC room events can be ignored safely while not affecting common
roster usage.
