RevisionTree.iter_changes(WT) should be as fast as WT.iter_changes(RevTree)

Right now our InterDirStateTree optimizer only thinks it is compatible if you use WT.iter_changes(other_tree), turning it around as other_tree.iter_changes(WT) goes via the generic (slow) path. On large trees, this can be particularly significant (30s vs 5s).

iter_changes is symmetric. So even if we can't easily change the _process_entry logic, we can just swap the trees, call iter_changes() then reverse the output.