R=go1.11.

The existing algorithm for type-checking interfaces breaks down in
complex cases of recursive types, e.g.:

	package issue21804

	type (
		_ interface{ m(B) }
		A interface{ a(D) }
		B interface{ A }
		C interface{ B }
		D interface{ C }
	)

	var _ A = C(nil)

The underlying problem is that the method set of C is computed by
following a chain of embedded interfaces at a point when the method
set for one of those interfaces is not yet complete. A more general
problem is largely avoided by topological sorting of interfaces
depending on their dependencies on embedded interfaces (but not
method parameters).

This change fixes this problem by fundamentally changing how
interface method sets are computed: Instead of determining them
based on recursively type-checking embedded interfaces, the new
approach computes the method sets of interfaces separately,
based on syntactic structure and name resolution; and using type-
checked types only when readily available (e.g., for local types
which can at best refer to themselves, and imported interfaces).

This approach avoids cyclic dependencies arising in the method
sets by separating the collection of embedded methods (which
fundamentally cannot have cycles in correct programs) and type-
checking of the method's signatures (which may have arbitrary
cycles).

As a consequence, type-checking interfaces can rely on the
pre-computed method sets which makes the code simpler: Type-
checking of embedded interface types doesn't have to happen
anymore during interface construction since we already have
all methods and now is delayed to the end of type-checking.
Also, the topological sort of global interfaces is not needed
anymore.

Fixes #18395.

Change-Id: I0f849ac9568e87a32c9c27bbf8fab0e2bac9ebb1
Reviewed-on: https://go-review.googlesource.com/79575Reviewed-by: Alan Donovan <adonovan@google.com>