Currently the implementation for calculating uses uses a HashSet to
collect all used packages. This leads to unstable behavior if the same
cleanup is applied multiple times.

This now uses a TreeSet instead to ensure lexicographic order is used.