make default value of rev keyword argument type stable

[JeffBezanson]

This lets us remove the method overwrite during sysimg build, and allows us to fix #9498. There's no user-visible change except that rev=nothing will now work (though it's undocumented), which seems harmless enough to me. We can easily remove this later if constant propagation permits.

[vtjnash]

We can easily remove this later if constant propagation permits

How well does #24362 do towards this eventual goal?

[JeffBezanson]

Just tried it; it works.

The change from branches to dispatch here is a small additional help since it allows more inlining.

[JeffBezanson]

Oops, I didn't measure carefully enough. I thought I was looking for allocations but actually the only issue here is dispatch time. Looking again, #24362 doesn't seem to help. I get this on that branch:

julia> f() = Base.Sort.ord(isless,identity,false,Base.Order.Forward)
f (generic function with 2 methods)

julia> @code_typed f()
CodeInfo(:(begin
        return $(Expr(:invoke, MethodInstance for ord(::typeof(isless), ::typeof(identity), ::Bool, ::Base.Order.ForwardOrdering), :($(QuoteNode(Base.Order.ord))), :(Main.isless), :(Main.identity), false, :($(QuoteNode(Base.Order.ForwardOrdering())))))
    end)) => Any

[JeffBezanson]

Aha, I suspect all that's needed there is more precise t-functions for and_int, or_int, and not_int on Bool.

[JeffBezanson]

Ok, that's still not quite enough. This is a case where we need to actually call or inline the code specialized for false.

[StefanKarpinski]

Specialization on default values would presumably also have a similar effect, no?

[vtjnash]

inline the code specialized for false
more precise t-functions

Ah, of course. The two next PRs I had planned. (noted briefly in the todo note in the initial PR).