Queues create async operations which are often processed faster than sync operations because queues force an async design.
You should compare Little's Law to a hard drive with and without native command queueing.
For example, you may have a classifieds site, where users will upload/attach images to a listing. This doesn't occur at a consistent rate, and you may have 5,10,20 workers or more you can spin up if it gets under heavier load.
Queues let you separate the workload of processing images to different optimized sizes, while allowing the rest of the system to operate normally. This isn't a situation where it's to work around back-pressure but to normalize and distribute load.
Much like distributed databases help in a lot of ways. If you are mostly read, then read-only replicas are likely the simplest answer. If you're write heavy then sharding may be your best bet. Need to distribute processing load that can be async, queues/workers are a good approach.
It's a tool, like any other... It doesn't replace distribution, replication or sharding, it's meant to be used in combination with.
Having a queue of messages allows optimization in ordering the messages.
This type of queueing can smooth out demand spikes at the cost of latency. The author's point is naively adding a queue doesn't solve scaling problems. Although, as you point out, it may enable optimizations that are otherwise impossible.
NCQ actively interrogates the current list of work and optimizes delivery around what needs to be done. that's not the op's point.
However, queues can "fix" an overload in one sense by making an engineering tradeoff of increased latency and additional complexity (e.g. SPOF[2]). Peak capacity handling didn't increase but overall server utilization can be maximized because jobs waiting in a queue will eventually soak up resources running at less than 100% capacity.
If response time (or job completion time) remains a fixed engineering constraint, queues will not magically solve that.
[1]http://en.wikipedia.org/wiki/Little%27s_law
[2]http://en.wikipedia.org/wiki/Single_point_of_failure