[GSB] Map invalid subject type to error type after resolving it

[theblixguy]

In #29491, we mapped an invalid subject type to an error type, but only mapping it if the subject type isn't a dependent member type is not enough - it's possible we have a dependent member type initially, but resolving it leads to some other type. So, we should delay the mapping instead. Otherwise, we could crash in scenarios like:

protocol P {
  associatedtype Assoc
}

enum SomeEnum {}

protocol Q: P where Assoc == SomeEnum {
  associatedtype Assoc: SomeEnum // crash
}

[theblixguy]

@swift-ci please test

[theblixguy]

@swift-ci please smoke test