Improvements compared to the previous implementation:
- avoid looking up the user document again, when the PAS plugin already did
  that job
- make it possible to call a single script when multiple sources of groups
  are based on the same documents, avoiding iterating unnecessarily on
  those same documents multiple times
- avoid repeating the same membership value (ex: when a user has multiple
  assignments with a common membership subset)
- avoid resolving the same relation more than once
- do not go from document value to relative URL only to go from relative
  URL back to document value at the next step
- move security group id extraction to unrestricted python, as the security
  overhead was taking a large amount of time
In a security setup with 8 scripts (all Assignment-based), 6 base
categories, and 4 Assignments (all valid), this implementation is 10 times
faster at producing the same group id set as the previous one.