doesn't know how to do LE, LT, GE, GT.  dict_richcompare can't do the
latter any faster than dict_compare can.  More importantly, for
cmp(dict1, dict2), Python *first* tries rich compares with EQ, LT, and
GT one at a time, even if the tp_compare slot is defined, and
dict_richcompare called dict_compare for the latter two because
it couldn't do them itself.  The result was a lot of wasted calls to
dict_compare.  Now dict_richcompare gives up at once the times Python
calls it with LT and GT from try_rich_to_3way_compare(), and dict_compare
is called only once (when Python gets around to trying the tp_compare
slot).
Continued mystery:  despite that this cut the number of calls to
dict_compare approximately in half in test_mutants.py, the latter still
runs amazingly slowly.  Running under the debugger doesn't show excessive
activity in the dict comparison code anymore, so I'm guessing the culprit
is somewhere else -- but where?  Perhaps in the element (key/value)
comparison code?  We clearly spend a lot of time figuring out how to
compare things.