mpf_sub bug (essentially same as reldiff bug)

[Torbjorn Granlund]

abbott at module.dima.unige.it writes:

  I wrongly get 0 if
     requested precision of a and b is 256
     a = 1   and b = 1+2^(-129)
     requested precision of rd is at most 64

  I get the RIGHT answer if
     a and b as before BUT
     requested precision of rd is increased beyond 64
     [I tried several values beyond 64, but not all of them :-]
  
Thanks.  There is no need to investigate the envelope of a bug;
a plain test case demonstrating the bug is sufficient.

This miscomputation bug has apparently been around since the first
release with mpf.  Not good.

--
Torbjörn