NaN behavior

postgres=# select 'NaN'::numeric = 'NaN'::numeric,
'NaN'::float8 = 'NaN'::float8;
?column? | ?column?
----------+----------
t | t
(1 row)

This behavior is inconsistent with most people's notion of "NaN" -- in
particular, it is inconsistent with IEEE754. I can understand why
Postgres behaves this way, and we probably can't easily change it (if we
want to continue indexing NaN values, that is), but I think it should at
least be discussed in the documentation.

Comments? I'll write up a doc patch, barring any objections.

-Neil