This also:
- adds an enum classification of Vararg types
- implements a few convenience functions
- switches some checks from jl_is_va_tuple to jl_va_tuple_kind(t) == JL_VARARG_UNBOUND
- updates jl_wrap_vararg to take two inputs
- displays such Varargs properly in the REPL

This leaves the old one in place and compares the results of the two.
This way we can catch bugs immediately, rather than trying to deduce
what went wrong from the indirect consequences.

This also makes tweaks to the algorithm to prevent matches like
Array{Tuple{Int}} <: Array{NTuple}

This also turns debugging output back on in preparation for the final implementation
The cmdlineargs test fails, but only due to the extra output generated.

Matching the symbols turns out to be really dangerous, it's necessary
to match pointers.

Needed for NTuple{Integer} declarations like found in sparsematrix.jl.

On current master, this gave true but
   Type{Tuple{}} <: Type{Tuple{Vararg}}
gave false. Now they both give true.

c.f. "Instantiate Tuple{Vararg{Int,3}} as Tuple{Int,Int,Int}"

this ensures that type-inference gets the same method signature from type-intersection
as will be used for the actual call