optimizer: propagate callsite inlining annotation across kwfunc

[aviatesk]

The callsite inlining annotations only works per-frame, but it might
appear unintuitive when applied to kwfuncs. This commits adds an extra
effort to propagate callsite inlining annotations across kwfunc abstraction:

julia> function someeval(@nospecialize(x); mod=Main)
           v = Core.eval(mod, :(x = $(esc(x)))) # by default this prevents inlining
           return v
       end
someeval (generic function with 1 method)

julia> let src = code_typed1((Any,)) do x
               someeval(x; mod=Main)
           end
           @test count(iscall((src, Core._expr)), src.code) == 0

           src = code_typed1((Any,)) do x
               @inline someeval(x; mod=Main)
           end
           @test count(iscall((src, Core._expr)), src.code) == 2
       end
Test Passed

[vtjnash]

This could be somewhat highly aggressive, since the user can include arbitrary code in the kwsorter (via default arguments), but seems fairly reasonable anyways

[vtjnash]

hm, wait, can't there be any number of callers, and this is just the first one that we saw, before we put it in the cache?

[vtjnash]

this seems like some very incorrect global state changes for the containing method?

[aviatesk]

yeah, somehow I assumed the containing method only has a single call of kwfunc at most, but it isn't true really.
The right way to support that is probably to propagate this information with a dedicated a lattice element.

[vtjnash]

This also comes up with positional arguments. For example:

f(a, b = (for _ in 1:10; for _ in 1:20; for _ in 1:30; end; end; end)) = a
g() = f(0)

So I am not sure anymore that we should do this.

[aviatesk]

This also comes up with positional arguments. For example:

f(a, b = (for _ in 1:10; for _ in 1:20; for _ in 1:30; end; end; end)) = a
g() = f(0)

So I am not sure anymore that we should do this.

yes, but for positional arguments we can enforce inlining by listing up all arguments (very verbose though), but for keyword arguments we currently don't have a way to enforce inlining.

[aviatesk]

Now I setup a new lattice element Kwfunc so that callsite inlining information on kwfunc call is propagated locally.
Though now this PR comes with some complexity, and I'm not too sure we really want to support this niche.