As suggested by David Chisnall. If I read the spec correctly, it would
be unsafe to use the release ordering for dec_ref_core: then the
following situation could happen:
```c++
// m_rc -> 2
// Thread 1:
unsigned x = atomic_fetch_sub_explicit(&m_rc, 1u, memory_order_release);
// Thread 2:
unsigned y = atomic_fetch_sub_explicit(&m_rc, 1u, memory_order_release);
// x = y = 1u, m_rc -> 2
```
That is, a release store operation is not required to be visible to
another release operation.
Herb Sutter also recommends the acq_rel ordering for reference counters.