Use depth-first scan for eliminating partial solutions instead of breadth-first
This reduces the peak number of lists in-memory from 2^k to k, and enables the
solver to eliminate most duplicates before it has instantiated the full set of
leaves.
Part of #857.