Instead of using the hierarchical nature of the ClipScrollTree to
calculate local clipping rectangles for primitives, do it when we are
about to render a particular primitive run. This will allow us to
handle custom clip chains while maintaining local clipping rectangle
optimizations. It also allows us to greatly simplify a lot of shader
code that used to deal with both clipping and scrolling nodes, as the
local clipping rectangle is now bundled into the primitives
local_clip_rect.

TransformOrOffset is the main struct that handles the efficient
transformation of clipping rectangles between compatible coordinate
systems. It will avoid doing expensive matrix operations as much as
possible, which should make these calculations cheaper -- important now
that they are done per-primitive run.