While working on adding line decorations, I kept running into
issues and extra complexity. Instead, I've refactored how
text shadow contexts are handled. The implementation is a bit
complex in parts, but conceptually this fits in much better with
how the frame builder, prim store and GPU cache work.

Now, text-shadow primitives add text runs and decorations as
sub-primitives, using an index buffer of primitives. This means:

* The code to manage the text-element stack and maintain correct
  paint order is much simpler.
* The GPU cache layouts are much simpler. Text runs and decorations
  within a text-shadow element have the same layout as normal
  primitives (no weird offsets and packed layouts).
* Text primitives that have both visual and shadows (which is the
  common case) share the same GPU primitive - so they only get
  built and uploaded once. This is quite a large performance win
  on pages with a lot of shadows.
* The code to truncate glyph positions is moved from the CPU
  to the vertex shader, which is a significant performance win.
* The paint order of shadows is reversed from the previous
  implementation, back to what it was originally.