Change extension types such that they must declare subtype relations to Object

[eernstg]

See the proposal in #3415. This PR changes the extension type feature specification such that the nullability of the representation type does not matter for the determination of the superinterface graph and subtype relationships of the types expressed based on a given extension type declaration:

An extension type is always a proper subtype of Object?, and as such it is potentially nullable (unless some further subtype relations exist). An extension type can declare that it implements T for various T, and this may cause the extension type to be a subtype of Object (directly or indirectly). In this case the extension type is non-nullable.

This change is a substantial simplification because the superinterface graph is shared among of all types E<T1 .. Tk> where E resolves to an extension type declaration (which includes the non-generic case when k is 0).

(In contrast, with the current feature specification, we could have an extension type E such that the superinterface graph of E<int> was different from the superinterface graph of E<int?>. They will now always have the same superinterface graph, though possibly with different actual type arguments in each node of the graph.)

[eernstg]

@scheglov, @johnniwinther, @chloestefantsova, the language team accepted this simplification of the rules about the superinterface graph structure of an extension type.

Sorry about the wasted work which has been done in order to handle the previous ruleset! Nevertheless, I'm sure we will be happy in the future about having these new, simpler rules.

[eernstg]

Is this one ready to land? The text still says that for the purpose of computing the Dart 1 least upper bound, we pretend that the superinterface graph has a top node. That's simple to say and simple to understand, but the existence of the top node may be controversial, even though it's clearly labeled as being only for this single purpose.