speed of mpz_odd_p and lazy allocation

[Torbjorn Granlund]

I think we should think twice before changing GMP's trickery about
expecting a one-limb alloation always.  This type of streamlining in GMP
saves many cycles when computing with small bignums.  I'd rather do more
micro-optimisation than less!

I think it might nevertheless be possible to allow lazy allocation using
some trickery.  Perhaps custom allocators could _mp_d could point to a
common scratch limb, for this situation?  This should create no problems
inside GMP, such as with threading, as read _mp_d[0] will never be
depended upon.

-- 
Torbjörn