I'm still flip-flopping on this one.  On the minus side, this change
risks breaking doctests without any enormous gain to show for it.

On the plus side, Eric Smith is currently working on implementing
sensible formatting for complex numbers (issue 1588), and again the
current form of complex formatting is going to cause (minor,
surmountable, but still ugly) difficulties for that implementation
when the empty presentation type is being used (e.g., format(3j, '')
or format(3j, '<20')).

I'll come up with a patch and see how much breaks in core Python.