WIP: max/min -> maxof/minof

[blakejohnson]

Implements the renaming of the binary max/min functions to maxof and minof as discussed in #2265. Now, if max or min are given a second argument, it is interpreted as the dimension or region over which to take the max or min.

Bye, max(a, (), 1)!

Note: this required updates in a lot of files. It passes all tests on my machine, but additional testing might be wise before merging.

[blakejohnson]

Whoops, forgot about extras/. I'll fix that in a minute.

[StefanKarpinski]

Woah. This is a surprisingly large change. That's not necessarily bad but it definitely needs to sit for a bit and collect some commentary.

[JeffBezanson]

The size of this change reflects that max is a very standard and commonly-used function as is. This is revealing.

[blakejohnson]

Well, we only need max/min on a dimension for non-scalar arguments. So, the other option is to make max(array, scalar) always mean max along a dimension, rather than max(append(vector, scalar))

[tshort]

I prefer that max([0,2,3], 2) gives [2,2,3] and leave operation along a dimension to another function or to Jeff's upgraded sweep_dim or whatever it would be called.

[StefanKarpinski]

See this is where it gets kind of ugly. Since max and min have this problem and they're in some sense related to mean and median (as statistics one can take of a set of values), it gets kind of weird that some statistics can take a dim argument and some can't. Using a dim keyword would solve this, of course.

[StefanKarpinski]

Unfortunately, there's a reason why this issue hasn't been satisfactorily resolved yet :-\

[blakejohnson]

@tshort I prefer that functionality in a statement like map(x->max(x,2), [0,2,3]).

[ViralBShah]

The code is actually a lot more readable now.

[blakejohnson]

@tshort Having ruminated on this for a day, I see now why your max([0,2,3], 2) => [2,2,3] is intuitive in some sense.