The current member constraint algorithm has a failure mode when it encounters a
variable `'0` and the only constraint is that `':static: '0`. In that case,
there are no upper bounds, or at least no non-trivial upper bounds.  As a
result, we are not able to rule out any of the choices, so if you have a
constraint like `'0 member ['a, 'b, 'static]`, where `'a` and `'b` are
unrelated, then the algorithm gets stuck as there is no 'least choice' from
that set.

The tweak in this commit changes the algorithm so that *if* there are no upper
bounds (and hence `'0` can get as large as desired without creating a region
check error), it will just pick `'static`. This should only occur in cases
where the data is flowing out from a `'static` value.

This change is probably *not* right for impl Trait in let bindings, but those
are challenging with member constraints anyway, and not currently supported.
Furthermore, this change is not needed in a polonius-like formulation, which
effectively permits "ad-hoc intersections" of lifetimes as the value for a
region, and hence could give a value like `'a ^ 'b` as the resulting lifetime.
Therefore I think there isn't forwards compat danger here. (famous last words?)