* Make InverseMapQuadToLocalSpace flatten transform to avoid 3D clipping
  glitches.

  While cc::MathUtil::InverseMapQuadToLocalSpace was documented as
  requiring callers to flatten the transform, only one of the four
  existing callers did this.  This changes it to flatten the transform
  internally.

  This fixes a bug where incorrect clipping rects were used in at least
  two of those three callers that were not flattening:  the callers in
  SkiaRenderer::CalculateRPDQParams and
  GLRenderer::UpdateRPDQWithSkiaFilters.  These callers were computing
  the minimal clip rect in local space (in a plane that might have a 3D
  transform) by transforming the clip rect (in the simple testcase, the
  window boundary), as a quad, into the local coordinate space,
  producing points that (with 3D) can have a nonzero Z components.  This
  nonzero Z component was then dropped.  However, the desired result was
  the point with a zero Z component in local space that would produce
  the correct X/Y position in device space with some arbitrary Z
  component in device space (that could then be dropped).

  This fixes the case tested in
  preserve3d-and-filter-no-perspective.html

* Don't try clipping at all when the transform has perspective, since
  when there is perspective, there could be areas on the local surface
  that are visible even though they are outside the bounds of the quad
  that maps to the device's clip.

  This fixes the case tested in
  preserve3d-and-filter-with-perspective.html

Bug: 923766
Change-Id: I5fbc679300ca52446552a1f5d90160309073d687
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2809095
Reviewed-by: vmpstr <vmpstr@chromium.org>
Commit-Queue: David Baron <dbaron@chromium.org>
Cr-Commit-Position: refs/heads/master@{#870500}