mpn_invert implementing ApproximateReciprocal.

[Torbjorn Granlund]

nisse at lysator.liu.se (Niels Möller) writes:

  Maybe you can use {sb|dc}pi1_divappr_q here? (But not the mu variant,
  since that would give a circular dependency...) According to the
  comments, these functions return a quotient which is correct or one unit
  too large, so you'd have to compensate for that.
  
we could speed up dcpi1_div_*.c by using mulmod_bnm1 instead of mul,
since we have remainder cancellation.  That should change thresholds
around in division quite a lot.

-- 
Torbjörn