Fix cycle detection for type aliases with wildcard arguments

[odersky]

#15507 escaped the cycle detector since when the right hand side was typechecked,
the type constructor _NestedSet2's type was [X] <: Any and the constructor was marked Provisional.
When faced with the application NestedSet2[?], the type checker tried to reduce the application, since
abstract types like NestedSet2 are not allowed to take wildcards (this would be equivalent in power to existential types, which are not supported in Scala 3). Reduction in a covariant context would give
Any, which means that the recursive constructor disappears.

There was a bug in safeDealias that caused this behavior also for non-variant variables. This got fixed in the 2nd commit.

Fixes #15507

[odersky]

How I fixed it:

This one was puzzling. I tried to instrument the checkNonCyclic function to see what went wrong. I did that for the original definition

type _NestedSet1[X] = Set[_NestedSet1[?]]

which did not produce an error and the variant

type _NestedSet1[X] = Set[_NestedSet1[Int]]

which did produce an error. But at first I could not spot anything, mostly because I had to print the types in toString form since showing them with the i interpolator produced cycles.

It was when I did a diff between the instrumentations of the two definitions that I saw that NestedSet1 had completely disappeared from the computed right hand side of the type. Instead the RHS was computed as Set[Any].

That caused the investigation to switch to Namer where the RHS was computed. Here, to be able to bootstrap we first assume a type [X] >: Nothing <: Any for NestedSet1 and mark the symbol as Provisional. This is arguably very dirty, but looked like the only practicable way to deal with recursive definitions like this. Then when we computed the RHS we had a NestedSet1 <: ([X] => Nothing .. Any) for the constructor and the TypeBounds Nothing .. Any for the argument. This is an illegal application since it would be equivalent to an existential type \exists X.NestedSet1[X] and Scala 3 does not have existential types. The typer will try in TypeApplications to reduce such existential types away and report an error in PostTyper if that does not work. In this case the reduction did succeed, we end up with just Nothing..Any.

Aside: In fact because of another problem the reduction did not give the TypeBounds type Nothing..Any but just the upper bound. That was fixed in [030e867](https://github.com/lampepfl/dotty/pull/15508/commits/030e867fae1b67469d5c5ba5bdd77c97f4a30ffb).

In either case the whole type was now just Set[Any], or after the fix Set[? >: Nothing <: Any]. NestedSet1 has completely disappeared so no cycle was detected.

The fix was then comparatively simple. If we have a provisional type constructor, we do not allow wildcard arguments. Either the constructor will become an alias type, then it's a cyclic reference error, or it will become an abstract type, then it's an irreducible application to wildcard arguments. In either case it's wrong.