Use consistent NaN behavior for max functions

[electrum]

NaN is effectively ignored, unless it is the only value. In other words, it is treated as the maximum value for min, and the minimum value for max.

From an implementation point of view, since it normally compares as the maximal value, we only need to special case it for maximum by reversing its comparator.

[dain]

Looks good.

[findepi]

Title talks about max, i see you;re changing greatest. Are you taking care of array_{min,max} and {min,max}_by as well?
after the change, Is min going to be produce same result as array_min(array_agg(...)) , and same result as reduce(array_agg(...), ... -> least(...)?

[electrum]

I consider greatest to be the scalar version of max. It should touch all of the max related functions, unless I missed any. Those should all produce the same result -- that's the goal of making these consistent.

Note that the NULL behavior is different, since array_min / array_max returns NULL if any element is null. The same is true for least / greatest.