reorder ml-matches to avoid catastrophic performance case

[vtjnash]

This ordering of the algorithm abandons the elegant insertion in favor of using another copy of Tarjan's SCC code. This enables us to abort the algorithm in O(k*n) time, instead of always running full O(n*n) time, where k is min(lim,n).

For example, to sort 1338 methods:

Before:
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 3, Base.get_world_counter());
  0.136609 seconds (22.74 k allocations: 1.104 MiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, -1, Base.get_world_counter());
  0.046280 seconds (9.95 k allocations: 497.453 KiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 30000, Base.get_world_counter());
  0.132588 seconds (22.73 k allocations: 1.103 MiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 30000, Base.get_world_counter());
  0.135912 seconds (22.73 k allocations: 1.103 MiB)

After:
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 3, Base.get_world_counter());
  0.001040 seconds (1.47 k allocations: 88.375 KiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, -1, Base.get_world_counter());
  0.039167 seconds (8.24 k allocations: 423.984 KiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 30000, Base.get_world_counter());
  0.081354 seconds (8.26 k allocations: 424.734 KiB)
julia> @time Base._methods_by_ftype(Tuple{typeof(Core.kwcall), NamedTuple, Any, Vararg{Any}}, 30000, Base.get_world_counter());
  0.080849 seconds (8.26 k allocations: 424.734 KiB)

And makes inference faster in rare cases (this particular example came up because the expression below occurs appears in @test macroexpansion), both before loading loading more packages, such as OmniPackage, and afterwards, where the cost is almost unchanged afterwards, versus increasing about 50x.

julia> f() = x(args...; kwargs...); @time @code_typed optimize=false f();
  0.143523 seconds (23.25 k allocations: 1.128 MiB, 99.96% compilation time) # before
  0.001172 seconds (1.86 k allocations: 108.656 KiB, 97.71% compilation time) # after

[vtjnash]

@nanosoldier runbenchmarks(!"scalar" && !"union", vs=":master")

[nanosoldier]

Your benchmark job has completed - possible performance regressions were detected. A full report can be found here.

[vtjnash]

As expected, this should be a big win for inference