fix(layout): give an anchor that spans pages the slice on each of them

[DemchaAV]

Problem

A LayoutAnchorNode wrapping content taller than a page emitted one resolved anchor per
page it occupied — and every one of them repeated the full measured subtree height.

A section spanning five pages, on a 320×150 page with a 20pt margin (content band 20..130):

page	reported y	reported height	reported top
0	20.000	453.250	473.250
1	20.000	453.250	473.250
2	20.000	453.250	473.250
3	20.000	453.250	473.250
4	91.150	453.250	544.400

Every top is hundreds of points off the page. Nothing could use an anchor on content taller
than one page: not "where does this start on page 2", not "how tall is it here". The
per-page occurrences were there, but each carried a number belonging to no page.

This was written down as a limitation rather than fixed, on the grounds that no consumer
needed the per-page extent yet. One does now, and the numbers turned out to be unusable
rather than merely imprecise.

Fix

Each occurrence reports the wrapped node's actual border-box slice on that page, taken
from the FragmentPlacement the compiler had already computed for it.

The placement was always the answer. A composite's fragments are emitted once per page it
occupies, and CompositeDecoration clamps each segment to that page's content band — the
same geometry that puts a spanning section's border on every page it crosses, per-page
margins included. LayoutAnchorDefinition was discarding it and reading
prepared.measureResult() instead. Reading the placement is also what keeps this one
formula rather than a second one drifting beside it.

Same case, after:

page	y	height	top
0	20.000	110.000	130.000
1	20.000	110.000	130.000
2	20.000	110.000	130.000
3	20.000	110.000	130.000
4	91.150	38.850	130.000

Checked against the thing it has to agree with

An anchored section and an accentLeft section of the same shape, laid out across the same
five pages — the border fragment is the independent implementation of "this box, clamped to
this page", so the anchor slices have to match it:

page	accentLeft (y, h)	anchor (y, h)	Δy	Δh
0	20.000, 110.000	20.000, 110.000	+0.000000	+0.000000
1	20.000, 110.000	20.000, 110.000	+0.000000	+0.000000
2	20.000, 110.000	20.000, 110.000	+0.000000	+0.000000
3	20.000, 110.000	20.000, 110.000	+0.000000	+0.000000
4	64.200, 65.800	64.200, 65.800	+0.000000	+0.000000

Page 4 includes the section's own 14pt bottom padding inside the border, and the anchor
reports it too.

Margins come off per edge

A margin is part of the flow extent the wrapper has to occupy but no part of the box being
reported, and a slice only meets the edges it actually contains:

• the top margin is inside the first page's slice;
• the bottom margin is inside the last page's slice;
• a middle page holds neither, so it is the whole content band;
• a node that fits on one page is both first and last, so it loses both — exactly as
  before;
• horizontally every slice spans the whole box, so left and right always come off.

Unchanged

• Anchors on content that fits on one page, which is every anchor that exists today.
  The atomic 8×8 marker, the marker carrying an asymmetric margin, and the marker inside a
  table cell all report the same numbers to 1e-9.
• Border-box semantics: margin excluded, padding included, the wrapped node's box and not
  its container's.
• One LayoutAnchorId across every slice — same groupKey, kind and index.
• No path or name inference anywhere.
• Public API: nothing added, removed or moved; extract-api --check is current
  without regenerating.

Verification

./mvnw -B -ntp clean verify -pl :graph-compose-core,:graph-compose-render-pdf,:graph-compose-render-docx,:graph-compose-render-pptx,:graph-compose-templates,:graph-compose-testing,:graph-compose-qa,:graph-compose-coverage -am
→ BUILD SUCCESS. javadoc:javadoc green; the knowledge gates green (surfaces current,
stability doc, claims, routes, bundle verify); the backends' no-op handler tests green.

ResolvedLayoutPassTest gains four cases, each written against a measurement rather than
against the implementation:

• a multi-page zero-margin anchor: first slice node-top → band bottom, last slice band top
  → node bottom, none of them the whole subtree;
• a middle page is the whole band;
• an asymmetric margin comes off the edge its slice contains, and horizontally off all of
  them;
• per-page margins: a page the rule widened to 40pt reports a 40..160 band, not the
  document default.

Proven fail-closed by putting measureResult() back: those four go red and the identity
case stays green, which is right — identity was never the broken part.

Lane: shared-engine — document.layout only. No Timeline file is touched; the word
does not appear in the diff.