This allows us to choose to rasterize any picture in an
arbitrary coordinate space.

This fixes a large number of real world pages that use 3d
transforms. It also simplifies fixing any remaining 3d
transform related issues.

The initial patch uses this to promote any picture that has
perspective into a rasterization root.

There are a number of benefits to this:
 * It simplifies a lot of the math involving plane splitting,
   since we don't need to deal with splits drawn in screen-space.
 * It makes caching perspective rendered pictures much easier,
   since we can cache them between frames even when the
   root coordinate system is animating the picture (this isn't
   implemented yet, but is a simple-ish follow up).

We can make use of this to render text shadows in local space,
as a performance optimization (which can then be cached between
frames and display lists).

Part of this change involves simplifying the plane splitting
code to work in a similar way to traditional brush images. It's
likely that in the future we can completely remove ps_split_composite
and draw plane splits with brush_image for better batching.

Introducing a new coordinate space allows for some more optimal
calculations of required raster rectangles, which reduces the
number of pixels that we draw for clip masks and transformed
pictures, in some cases.

Follow ups:
 * Support arbitrary rasterization scales for raster roots, for
   better quality / performance.
 * Support caching of pictures as discussed above.

The split composite shader is now quite similar to a normal
brush shader (uses prim header, supports normal clip masks
etc). The only real difference is that it draws a quad polygon
from local points instead of a rectangle. It should be easy
to unify this into the brush_image shader in the future.

Add a temporary "ClipNodeCollector" interface to correctly
handle clips that are ancestors of the current rasterization
root. Although it's not enforced by the API, I have verified
that any time a clip from a spatial node that is an ancestor
of the current rasterization root is applied on all primitives
that are part of that picture. This means we can safely apply
those clips during compositing of the rasterization root rather
than on each individual primitive. This fixes correctness
issues with clips across rasterization roots, and is also a
quite significant performance win in some cases, since the
clip mask is now applied once when the rasterization root
is drawn, rather than on each indidvidual primitive.

The end result of this is actually quite good - it will allow
us to simplify some of the clipping code in the future, and is
a performance win. The implementation leaves a lot to be
desired, but gets the required result without having to fix
up both gecko and wr clipping API mismatches. The long term
solution for this is to update gecko and WR to make better use
of clip chains on stacking contexts, rather than putting
clips from the stacking context on each individual primitive (we
have discussed doing this previously, but it's low-ish priority
for now given other bugs).