Unwind a hack whose stated purpose was to register a potential member
edge from an extension to the extended type. In reality, this only
registered a plain member dependency on 'deinit'. This edge is
insufficient in isolation to cause a rebuild of a dependent file in the
case where a type and its extension live in separate files. However, we
appear to have been saved by the redundancy in edge registration because the
lookup for the extended type will register a top-level or nominal
dependency (for an unqualified or qualified reference respectively). The
worry there is if a protocol conformance edge *should* flip a previously
private nominal dependency edge to a cascading edge. In such a case, the
old code would not have been able to make the cascading edge promotion,
and we would have potentially miscompiled by not rescheduling dependent
jobs.
