Graph.find_differences() is confused by shortcuts

At the moment, find_differences() implements a naive search that can easily get confused by shortcuts.

For example, with:
# NULL_REVISION
# |
# a
# |
# b
# |
# c
# |
# d
# |\
# e f
# | |\
# i | h
# |\| |
# | g |
# | | |
# | j |
# | | |
# | k |
# | | |
# | l |
# |/|/
# m n

It will find 'k' as a common node, and won't continue and find that 'f' is actually not unique to 'n', but is actually also common.

There is work which has been done towards this end, but at the moment it does not perform well on the bzr graph (presumably it is too pessimistic about when it can stop searching common nodes.)
this is slightly related to bug #165307.

I will associate my branch with this bug.