Prevent duplicate relation count queries

[ydah]

This follows up on #5708, which prevents concurrent Fibers from loading the same relation page more than once.

RelationConnection#limited_nodes remains independently memoized outside that lock. With last: pagination and no existing relation limit, it calls relation_count(sliced_nodes). If that query yields, another Fiber can enter the same @limited_nodes ||= ... block before the first one finishes.

Under AsyncDataloader, this can happen when connection fields resolve separately:

• edges or nodes calls load_nodes
• pageInfo.hasPreviousPage calls limited_nodes directly

Both Fibers can therefore issue the same COUNT query.

connection = Connection.new(
  relation,
  last: 2,
  context: { dataloader: GraphQL::Dataloader::AsyncDataloader.new },
)

Sync do
  [
    Async { connection.nodes },
    Async { connection.has_previous_page },
  ].each(&:wait)
end

Before this change, the relation is counted twice. After this change, it is counted once.

This PR protects both limited_nodes construction and the final node load with the same Monitor. Monitor is reentrant, so load_nodes can safely call limited_nodes while holding it without risking ThreadError: deadlock; recursive locking.

The synchronization is only used with AsyncDataloader; other dataloaders retain the existing unsynchronized path and its current overhead.

[rmosolgo]

LGTM, but I think it needs one change now from a fix on master.

I haven't seen Monitor before -- what's the advantage over Mutex?