Convert ShenandoahOldGeneration to use Atomic<T>, this also fix the code issue below(missing volatile may not causing any issue as long as AtomicAccess is always used to access the fields)

>>> orignal description

Came across the source code of ShenandoahOldGeneration, and noticed that we use AtomicAccess to read/write to fileds `_promoted_expended`, `_promotion_failure_count`, `_promotion_failure_words` but non of them is declared with volatile modifier.

It supposed to be a small fix by just adding the missing `volatile` modifier, but instead of adding `volatile` modifier, I think we cloud use Atomic<T> instead of static APIs from AtomicAccess.

```
  // Bytes of old-gen memory expended on promotions. This may be modified concurrently
  // by mutators and gc workers when promotion LABs are retired during evacuation. It
  // is therefore always accessed through atomic operations. This is increased when a
  // PLAB is allocated for promotions. The value is decreased by the amount of memory
  // remaining in a PLAB when it is retired.
  size_t _promoted_expended;

...

  // Keep track of the number and size of promotions that failed. Perhaps we should use this to increase
  // the size of the old generation for the next collection cycle.
  size_t _promotion_failure_count;
  size_t _promotion_failure_words;

```


```
void ShenandoahOldGeneration::reset_promoted_expended() {
  shenandoah_assert_heaplocked_or_safepoint();
  AtomicAccess::store(&_promoted_expended, static_cast<size_t>(0));
  AtomicAccess::store(&_promotion_failure_count, static_cast<size_t>(0));
  AtomicAccess::store(&_promotion_failure_words, static_cast<size_t>(0));
}

...

void ShenandoahOldGeneration::handle_failed_promotion(Thread* thread, size_t size) {
  AtomicAccess::inc(&_promotion_failure_count);
  AtomicAccess::add(&_promotion_failure_words, size);

  LogTarget(Debug, gc, plab) lt;
  LogStream ls(lt);
  if (lt.is_enabled()) {
    log_failed_promotion(ls, thread, size);
  }
}


```