For caching staged methods, don't restrict the number of varargs to match

[timholy]

Note this is a PR against kf/staged, not master.

This fixes the main remaining issues with splatting:

julia> stagedfunction h(a...)
           @show a
           :nothing
       end
h (generic function with 1 method)

julia> h(1:5, 3, 3:3)
a => (UnitRange{Int64},Int64,UnitRange{Int64})

julia> h(1:5, 3, 3:3)

julia> h(1:5, 3:3, 3)
a => (UnitRange{Int64},UnitRange{Int64},Int64)

julia> h(1:5, 3:3, 3)

julia> h(1:5, 3)   # not correct

julia> h(3, 1:5)
a => (Int64,UnitRange{Int64})

I suspect that one remaining problem is due to this line. What's the reason for allowing a match when n == lensig-1?

[timholy]

That relatively complicated conditional can apparently be simplified as in my second commit. But it seems that n+1 is important; if I change it to n and try a clean build, I get this:

...
inference.jl
osutils.jl
Killed
*** This error is usually fixed by running 'make clean'. If the error persists, try 'make cleanall'. ***

Interestingly, if I let the build go through with n+1 and then change it to n and just rebuild libjulia.so, the tests mostly work (until we get to the staged.jl test). Is this for some special case in inference bootstrapping?

[Keno]

I took these commits when I rebased #7474, so closing this in favor of that.

[timholy]

Great, thanks.