This is a re-land of CL:2780489, but this time also remove a bogus
expectation file (added in CL:2071386), to avoid Mac failures.
Original commit message text follows:

We used to apply the whole stitched LayoutBox clip rectangle to every
fragment (we don't store clip rectangles per fragment). To fix this,
post-process the clip rectangle, to clip it against the bounds of the
fragment.

Fixes one test, but test coverage seemed low, so I added 3 more, one for
each writing mode.

This also fixes the following tests when CompositeAfterPaint +
LayoutNGBlockFragmentation are enabled:

  external/wpt/css/css-multicol/composited-under-clip-under-multicol.html
  fast/multicol/composited-relpos-clipped.html
  fast/multicol/composited-relpos-in-clipped.html

.. but note that the expectation files for the latter two are wrong
(pre-CompositeAfterPaint behavior is expected), so they will show up as
failing still.

Long-term we should probably consider storing a clip rectangle per
fragment. There are callers of the original
NGPhysicalBoxFragment::OverflowClipRect() (the one that doesn't take an
incoming break token) in NGBoxFragmentPainter that still don't do the
right thing (no access to the incoming break token there, so no easy way
to fix).

Bug: 1191162, 829028
Change-Id: If0d9e2c636d553186deb8cc3b022d8f729b3f1be
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/2809397
Reviewed-by: Koji Ishii <kojii@chromium.org>
Commit-Queue: Morten Stenshorne <mstensho@chromium.org>
Cr-Commit-Position: refs/heads/master@{#869946}