This PR removes the window variable in copyto_stencil_kernel!, and instead, uses different windows depending on gid.

I'm actually not sure how the gpu even allows this block of code:

            (v, i, j, h) = cart_ind((Nv, Nq, Nq, Nh), gid).I
            hidx = (i, j, h)
            idx = v - 1 + li
            window =
                idx < lw ?
                LeftBoundaryWindow{Spaces.left_boundary_name(space)}() :
                (
                    idx > rw ?
                    RightBoundaryWindow{Spaces.right_boundary_name(space)}() :
                    Interior()
                )
            setidx!(
                space,
                out,
                idx,
                hidx,
                Operators.getidx(space, bc, window, idx, hidx),
            )

because

• the type of window depends on the value of idx
• the value of idx depends on the value of v
• the value of v depends on the value of gid
• the value of gid depends on the individual thread, which is most definitely non-deterministic

The only way I could see this being type stable is if the compiler can determine that all three branches result in the same return type, and that, then, it doesn't matter what the value of window will be within the scope of copyto_stencil_kernel!.

This is surprising to me. Any insights @vchuravy?