Yes, these are both good approaches. I also came to the conclusion that, in general, there is no escaping writing C code (or something like it) for launching asynchronous tasks.

As for ParallelSubmit etc., I did get an example working with LocalSubmit, though it took a bit of finagling. You can find it under the scope section here: https://resources.wolframcloud.com/PacletRepository/resources/ChristopherWolfram/WebServer

I also had problems killing the LocalSubmitted task, or getting it to share variables with the main kernel, which sort of defeats the purpose.

Great post and paclet! Glad someone else found my code useful! I know how to make the function that will not block the kernel.

1. To do this you would have to create an asynchronous task directly on the C side, but then the very idea of writing a server only in a top-level language is already broken.
2. You could just make the server run by default on a parallel kernel with ParallelSubmit/LinkCreate and return a task that can be stop the server if it necessary.