Add Pade approximation for exp. From Richard Fateman.

Experiments show that it is 25% faster than exp-qd/reduce.  However,
it loses 3 bits of accuracy compared with exp-qd/reduce.