I think ideally all pm would be controlled by the runtime_pm and qos frameworks. You then don't need to define explicit power states like 'sleep' or 'suspend', you instead simply use what you need for a given performance spec and naturally use a minimum of power. I think that is the ultimate plan

It's not that simple. There's a lot of predicting the future when it comes to power management. When you put things to sleep in order to save power you add latency, because waking up the hardware and restoring its state takes a long time. It's all tradeoffs, and "use minimum power" is taking the slider and shifting it all the way to one side.