Hang in the test-case from #1127 with extra assertions.

[emilio]

With the fix for the panic in #1128, we get through it and generate code correctly, but on llvm 4.0 we generate a cyclic alias reference, so we get stuck resolving items in assert_every_item_in_a_module (enabled only with extra assertions).

The test is:

// bindgen-flags: -- -std=c++11

template <typename> struct A;
template <typename, typename, bool> struct _Map_base;
template <typename _Pair, typename _Traits>
struct _Map_base<_Pair, _Traits, true> {
  using __hashtable_base = A<_Traits>;
  using key_type = typename __hashtable_base::key_type;
  using mapped_type = typename _Pair::type;
  mapped_type &at(const key_type &);
};
template <typename _Pair, typename _Traits>
auto _Map_base<_Pair, _Traits, true>::at(const key_type &) -> mapped_type & {}

The method definition does surprisingly matter, so there's something fishy going on there, but I'm punting on landing the test-case for now because the fix fixes the issue, and also unblocks updating bindgen in mozilla-central.

[emilio]

Bisect points to c736faa, which is where the assertion was introduced, so that's not really helpful... :)

[fitzgen]

We have debug-only sets that we use to assert against infinite loops elsewhere (item ancestors iirc) so we should use them here as well.

Of course, the bigger issue is that we shouldn't be generating cyclic aliases at all...