I also want what you're describing. It seems like the ideal "data-in-out" pipeline for purely compute based shaders.
I've brought it up several times when talking with folks who work down in the chip level for optimizing these operations and all I can say is, there are a lot of unforeseen complications to what we're suggesting.
It's not that we can't have a GPU that does these things, it's apparently more of a combination of previous and current architectural decisions that don't want that. For instance, an nVidia GPU is focused on providing the hardware optimizations necessary to do either LLM compute or graphics acceleration, both essentially proprietary technologies.
The proprietariness isn't why it's obtuse though, you can make a chip go super-duper fast for specific tasks, or more general for all kinds of tasks. Somewhere, folks are making a tradeoff of backwards compatibility and supporting new hardware accelerated tasks.
Neither of these are "general purpose compute and data flow" focuses. As such, you get the GPU that only sorta is configurable for what you want to do. Which in my opinion explains your "GPU programming seems to be both super low level, but also high level" comment.
That's been my experience. I still think what you're suggesting is a great idea and would make GPU's a more open compute platform for a wider variety of tasks, while also simplifying things a lot.