Alter scale selection for NUMERIC division and transcendental functions

so that precision of result is always at least as good as you'd get from
float8 arithmetic (ie, always at least 16 digits of accuracy).  Per
pg_hackers discussion a few days ago.