add Order.rev() for reverse sorting

[aplavin]

Sorting with by=x -> (x[1], x[2]) sorts by x[1] and then (for x[1] duplicates) by x[2].

Then, sort(X; by) sorts in the ascending order of x[1] and x[2], sort(X; by, rev=true) sorts in the descending order of both.

But there's no way to sort in ascending order of x[1] and then descending order of x[2]. This PR makes it possible, with by=x -> (x[1], Order.rev(x[2]).

What do you think of this interface? I imagine this Order.rev function being public, but not exported.

[nsajko]

Couldn't you base this on Ordering somehow?

[aplavin]

Idk, what do you have in mind? ReverseOrdering exists but it basically serves rev=true.

[LilithHafner]

I think this case is handled adequately by the lt ordering. lt = (a,b) -> isless(a[1], b[1]) || isequal(a[1], b[1]) && isless(b[2], a[2]). Yes, Order.rev would be concise, but I think that it is better to use functions to express complex comparisons rather than introduce additional API.

Having optimizations that dispatch on by ordering that return objects with Order.rev components could shift that balance a bit, but I don't see those optimizations being worth the complexity they would require.

[aplavin]

Sure, it's possible to express any kind of sorting is lt already, and by argument is not necessary in general. It's just not too readable and more error-prone: eg, compare lt = (a,b) -> isless(a[1], b[1]) || isequal(a[1], b[1]) && isless(b[2], a[2]) and lt = (a,b) -> isless(a[1], b[1]) || isequal(a[1], b[1]) && isless(a[2], b[2]). Is it easy to spot the difference? To make a typo? Much cleaner to have by=x -> (x[1], Order.rev(x[2])) and by=x -> (x[1], x[2])!
And this is just two-part key, think about three or more.