feat(svm): clip the drawing to what the metafile clips it to

[andiwand]

🤖 Generated with Claude Code

Stage 5 of #772, stacked on #786 → #785 → #784 → #779 — review those first;
this branch's base is feat/svm-bitmaps.

Clipping was the largest gap left by occurrence: of 1125 metafiles harvested
from the fixtures, 1124 intersect a clip rectangle and 50 intersect a
region. All of it was skipped, so anything a metafile drew outside its own clip
was shown in full.

Intersecting clips are nested groups

An svg clip-path names one shape, and ISECT… intersects, so two clips
become two groups, one inside the other. The state's clip is therefore a stack
of shapes rather than one: ensure_clip keeps the groups the next drawing
action still wants, closes the ones it does not, and opens what is missing.
That also makes a POP restoring the clip nothing special — the next action
closes what it has to, and the document's last groups close before </svg>.

Reading a region

A region streams as a band list (RegionBand::load) — a union of rectangles —
and, from version 2, as the poly-polygon those bands were rasterised from. The
polygons are the better outline where they are there. Both end up in one
<path> with clip-rule="evenodd": bands are disjoint, so union and even-odd
agree, and a poly-polygon wants even-odd anyway.

Two details worth naming:

• vcl does not re-scale a clip when the map mode changes — SetMapMode's
  own comment says "clip regions are not re-scaled" — so the shape is
  transformed once, where the action sets it.
• A file sets its drawing area and then intersects the region of the same
  rectangle. That is one group, not two; intersecting a clip with the shape
  it already has is skipped.

MOVECLIPREGION stays unimplemented: it occurs nowhere in the corpus and would
have to move path data that is already written out.

Verification

4 new tests (29 in the svm suite): a clip wraps what follows, two clips nest,
a pop closes the group its push drew in, and the same clip twice is one group.

Rendered every changed page against the reference with compare-html --driver chrome: 1388 matched, 15 differ, and every one of those differs by 0.001% to
0.03% of its pixels — the antialiased edge where a clip now trims a hairline
at the drawing's border. Cropping the largest of them side by side shows two
identical charts. No text changed anywhere.

Reference output regenerated (19 private, 4 public) and the pin advanced in
this branch.