use named tuples as keyword varargs. fixes #4916

[JeffBezanson]

Same highly-biased benchmark as #21915 (~100x speedup):

julia> f(;kw...) = g(;kw...);

julia> g(;kw...) = kw;

julia> c = g(a=1,b=2);

julia> @time for i = 1:10000; f(a=1,b=2); end
  0.000010 seconds

julia> @time for i = 1:10000; f(;c...); end
  0.000577 seconds (20.00 k allocations: 625.000 KiB)

This also helps a lot towards #9551. The PR makes the lowering of keyword argument methods simpler and easier to optimize (note the net code deletion). In fact I think just a bit of constant propagation (#24362) on top of this will solve the problem entirely.

Also fixes #9972.

[quinnj]

I wonder if it'd be worth mentioning that keyword arguments can no longer be repeated; I know that wasn't changed explicitly in this PR, but seems like a good place to mention it anyway.

[vtjnash]

What constant propagation would that be? Currently that PR can’t really do anything for NamedTuples (which would likely need more of a MemSSA or GVN pass).

[JeffBezanson]

The code here calls haskey to see whether a keyword argument was passed or should get its default value. If we can remove the branch then code with keyword arguments can be fully inferred. In fact it already works in many cases, if the passed value has the same type as the default value.

[ChrisRackauckas]

What's the timing now that #24362 merged?

[yuyichao]

It's not measuring anything...

julia> f(;kw...) = g(;kw...)
f (generic function with 1 method)

julia> g(;kw...) = kw
g (generic function with 1 method)

julia> const c = g(a=1, b=2)
(a = 1, b = 2)

julia> @time for i in 1:1000; f(a=1, b=2); end
  0.000003 seconds

julia> @time for i in 1:1000; f(;c...); end
  0.000003 seconds