Optimize memory orders for weakly ordered architectures.

In `RefCount::Unref()`, `fetch_sub()` with `memory_order_release` rather than
`memory_order_acq_rel`. Then, only in the unexpected path when this was the last
decrement (due to `IsUnique()` called before, this can happen only if another
thread calls `Unref()` at the same time), `atomic_thread_fence()` with
`memory_order_acquire`.

This removes the unnecessary acquire semantics from the expected path.

In `StableDependencyDefault::EnsureAllocatedSlow()`, `compare_exchange_strong()`
with a separate `memory_order_release` for the expected success path and
`memory_order_acquire` for the unexpected failure path (which can happen only
another thread calls `EnsureAllocatd()` at the same time), instead of using
`memory_order_acq_rel` for both.

In the unexpected path there is no change because `compare_exchange_strong()`
implicitly demotes `memory_order_acq_rel` to `memory_order_acquire` there.
This removes the unnecessary acquire semantics from the expected path.

Cosmetics: in `RefCount::{Ref,Unref}()`, replace `if` with `if constexpr`,
and split a complex conditional into separate statements.
PiperOrigin-RevId: 915941386