* New function: mpf_cmp_2exp.  Good for termination conditions in Taylor
  series.

* Allocate PREC+1 limbs as soon as PREC is set or changed for a float.
  That eliminates the need for mpf_realloc calls that otherwise would
  clutter the code.

* We should probably handle truncatation in more cases.  In trivial
  cases we currently just return, even if res->prec << res->size.

* Strings (in mpX_get_str) and bignum objects should not be allocated using
  the same function.

* Exactly how to round stuff is a problem.  We allow prec+1 limbs now when
  we can't predict the number of limbs.  This has a major disadvantage:
  set(x,x) might change the value of x.
