That's not how I understood it at all. I always assumed that it just promoted explicitly acknowledging when the effect of an async block would be present, but that the normal yield points (traditionally syscalls and IO requests in OS threads) generally still held for the actual execution of the other code.
I mean, if we aren't allowing separate code paths to execute (and thus reducing wait on resources such as IO), what's the point?
Put it this way: the language implementation won't preempt you except at your explicit yield points (some standard library functions for things like I/O will and should be yield points that you have to call with async/await). If the language implementation happens to be running as a userspace process on a preemptive multitasking OS then it will still be subject to the same preemption rules as any other userspace process on that OS, but if anything this is usually counterproductive (e.g. priority inversions are almost guaranteed); recommended practice when working in green-thread-based systems is to run with one thread per CPU core and maybe even pin threads to cores if the OS lets you do that, because the language's own M:N scheduling will keep those threads fully occupied and OS scheduling is only going to get in the way.
I find that having one implicit, pervasive, uncontrolled effect in a given piece of code is ok, but having multiple uncontrolled effects is very much not ok, because the ways those effects will interact can be very surprising. E.g. https://glyph.twistedmatrix.com/2014/02/unyielding.html gives an argument for using explicit async/await rather than go-style implicit (green) threads, in terms of how task switching will interact with state mutation. You can make similar arguments in terms of how task switching interacts with error handling or logging or database transactions or... - one uncontrolled effect is fine, multiple uncontrolled effects cause chaos when they interact. Unfortunately the kind of examples where one can see multiple effects in a single application tend to be quite big by their very nature (and if your app is small enough to only need to have one kind of effect, then using a language in which that particular effect is uncontrolled is probably fine, perhaps even a good idea).