NaN divided by zero should yield NaN

Dean Rasheed questioned this longstanding behavior:

regression=# SELECT 'nan'::float8 / '0'::float8;
ERROR: division by zero

After a bit of research I think he's right: per IEEE 754 this should
yield NaN, not an error. Accordingly I propose the attached patch.
This is probably not something to back-patch, though.

One thing that's not very clear to me is which of these spellings
is preferable:

if (unlikely(val2 == 0.0) && !isnan(val1))
if (unlikely(val2 == 0.0 && !isnan(val1)))

I think we can reject this variant:

if (unlikely(val2 == 0.0) && unlikely(!isnan(val1)))

since actually the second condition *is* pretty likely.
But I don't know which of the first two would give better
code. Andres, any thoughts?

regards, tom lane