Add a proxy backend for multi-GPU copies
Right now multi-GPU copies happen directly in the DRM backend. Apart from making the code more complex and being the only reason why we still need to perform rendering in the backend, the main downside is that compositors can't easily disable the multi-GPU copies if they want to. In other words, compositors can't perform composition with the destination GPU, because the DRM backend will always perform a multi-GPU copy.
We could de-couple the multi-GPU copy from the backend, for instance by introducing a new "proxy backend" that performs multi-GPU copies for output buffers.