This commit reworks how drop-shadows are handled. We now add two
primitives, each one referencing the same Picture.

The first primitive draws the drop shadow by blurring the alpha
channel of the picture, and generating screen-space UVs after
applying a local offset.

The second primitive draws the main image with a normal
brush_image shader blit.

There are a couple of hacky bits of code to make this work, but
it's such an improvement on the current state of drop-shadows
that I think it's worth getting this merged and tidying those
up later (mentioned as TODO comments in the patch).

Also, finally remove the hardware composite shader, since this
removes the last use of it.

Fixes #2584.
Fixes #2374.

1) In picture.rs, the wrong GPU cache handle was being invalidated.
   This was causing visual glitches when scrolling pages with
   drop-shadows present.

2) Fix an edge case that was occurring when the visual content of
   a drop-shadow was scrolled off-screen, but the bounding rect of
   the drop-shadow itself was still partially visible. In this case
   the picture would not get prepare_prim_for_render() called, but
   the drop-shadow brush image would still attempt to draw, which
   would reference an invalid render task in the render task tree.
   Correctly drawing enough of the off-screen visual contents to
   allow correct drawing of the partial shadow will be done as a
   follow up - for now, we will just stop drawing the shadow if
   the main visual contents becomes invisible, which can result
   in a noticeable pop of the shadow in some cases, but is still
   a big improvement on current drop-shadow support. To do this,
   the logic in batch.rs is inverted. Instead of checking for
   an intermediate surface, and then the filter mode, the code
   now checks the filter mode first, followed by the surface. This
   means that in the case above, we can detect that the visual
   contents of the drop shadow were off-screen, and skip trying
   to draw the alpha mask of that as a shadow.