fix #31703, type intersection bug with chains of vars in invariant position

[JeffBezanson]

The problem here was the rule for intersecting var T inside e.g. Ref{T} ⊓ Ref{Foo{S}}. We were intersecting Foo{S} with the upper bound of T, and then (sometimes) doing a side check that that type is consistent. (The "sometimes" was added in #29380, basically as a different attempt to fix this same problem.) But that is not correct: T needs to equal Foo{S}, so Foo{S} must be fully within its bounds. The catch is that that doesn't need to hold for all values of S though; we are free to restrict S while forming the intersection. So here I changed the algorithm to always check T.lb <: Foo{S} <: T.ub, but with all variables treated as existential. Seems to work. This subtype_in_env_existential trick might have other uses in intersection as well.

Fixes #31703

[vtjnash]

Is this related to #29524 (see reduction near the end) which had:

julia> typeintersect(Ref{Tuple{T}} where T, Ref{_1} where _1<:Tuple{Real})
Union{}

[JeffBezanson]

Wow, yes, it does seem to fix that case too.