Refactor runner declarations should use the submodule system, to allow for greater flexibility

[MaeIsBad]

While experimenting with the newly merged amdgpu native contexts using microvm I've found that microvm is quite inflexible for users who want to modify it's behavior. In this case I found that in order to prevent microvm from setting -nographic I had to set graphics.enable = true, which causes microvm to set -device "virtio-vga-gl", which is incompatible with -device virtio-gpu-gl.

Using the qemu runner as an example, but this would also apply to others.

Instead of the current setup, where the qemu runner declaration is one big function that creates a command from microvm configuration options, it might be sensible to instead have the qemu runner expose a "microvm.qemu.args" option(the same as the existing extraArgs) as well as a readonly command option, with other modules using the option to configure qemu.

This would already allow users to remove any problematic arguments by doing a hack, like:

      options.microvm.qemu.args = lib.mkOption { apply = lib.subtractLists [ "-device virtio-vga-gl" ]; };

but the real improvements would be in the ability to further create helper options, like for example microvm.qemu.devices.*, allowing overwriting the options configured for any device.

I'd be up for implementing the changes if @astro would be willing to merge them

[astro]

I have been contemplating the same idea but I cannot find time to help.

There is one fringe use-case I would like to keep support for: extendModules on a finalized NixOS config with a host-supplied nixpkgs input as done in pkgs/build-microvm.nix. It doesn't have to be API-compatible but it allows more secure provisioning.

[SuperSandro2000]

This would also allow us to use assertions instead of throws and collect all errors at once and not one after the other.