fix #38888, pessimistic sparam inference with concrete upper bound

[JeffBezanson]

The problem here is that we infer the constructor with signature T(x::S) where S<:Int, but that's type-equal to T(x::Int), so it can be picked up from the cache later. But that specialization doesn't work when called, since it thinks it doesn't know the value of S. I think the simplest fix is to do a pass refining unknown static parameters to their upper bounds when possible (concrete, only occurs covariantly). I also tried an alternate fix where we avoid inferring T(x::S) where S<:Int by normalizing the signature in specialize_method. That also works but seems more complex. We might want to do it in addition though.

I also noticed a case where we seem to return Nothing instead of nothing (!!)

[Keno]

Doesn't this just punt the problem one level down the line, as in:

struct S38888{T}
    S38888(x::S) where {S2<:Int,S<:S2} = new{S}()
end
f38888() = S38888(Base.inferencebarrier(3))
@test f38888() isa S38888

[JeffBezanson]

Hmm, I guess in that case the other fix I described is more comprehensive.

[KristofferC]

Ready to go?