Split display/window wrappers

[ids1024]

WIP implementation of #37. Currently compiles on Linux, and splits parts of Wayland code to not need duplication for multiple windows.

Seems best to start with this API change before moving on to #40.

[notgull]

Context is expected to be created once per connection, which brings up how often Surfaces are expected to be created. Should it be created once per surface, kept around for the lifetime of that surface, and dropped only when the surface is dropped? Or would it be perfectly cheap to create and drop Surfaces, like so:

match event {
    Event::RedrawRequested => {
        let surface = Surface::new(&context, my_window);
        /* .. */
        drop(surface);
    }
}

If the latter, we should move some of the allocations out of Surface into the Context. In either case, this should be documented.

[ids1024]

I'm assuming one would keep the Surface and continue using it. So an allocations that are specific to the surface would be there.

[notgull]

I'm wondering if it would be more idiomatic to make set_buffer take &Context as an argument. This would mean that we wouldn't need to have an Arc<Context> for Wayland and X11. This is less of a performance thing and more of an "is this more idiomatic?" thing.

LGTM otherwise

[ids1024]

Maybe. Though taking a &Context could be a bit more inconvenient after #40 since that will add more methods to Surface. And then you either have to pass it to all of them, or it's necessary to determine which methods might need something from the context (on some backends).

There could also be weird behavior if you passed a different context, though that probably isn't too much of an issue.

(Perhaps the ideal solution would be a lifetime bound on Surface if Rust had some kind of magical solution to make self-referential structs convenient, but that doesn't exist and makes lifetime bounds often awkward to deal with.)