refactor(engine): extract the layer-stack placement into its own compiler

[DemchaAV]

Why

Continues the LayoutCompiler breakdown. compileStackedLayer (~134 lines — admit the
stack band, place each layer in z-order, emit fill/overlay decoration, record the node) is
the next-largest compile* method. Unlike compileSplittableLeaf (#360, self-contained),
its per-layer placement recurses back through the compiler (placeStackLayer →
compileNodeInFixedSlot), so this is a method object with a host callback, not a
standalone unit — it trims LayoutCompiler without fully decoupling the recursion.

What

• Moved compileStackedLayer into a package-private StackedLayerCompiler.compile(host, …);
  LayoutCompiler dispatches to it and drops the private method (1196 → 1062 LOC).
• The collaborator calls back host.placeStackLayer(…) / host.admitAtomicBlock(…) (both now
  package-private) — placeStackLayer stays in LayoutCompiler because the STACK branch of
  compileNodeInFixedSlot also uses it.
• Introduced a CompileContext record bundling the five cursor/emission params
  (state / prepareContext / fragmentContext / nodes / fragments) so the collaborator
  takes one context argument; dropped the unused availableWidth param; repointed a {@link}
  and a comment.

Tests

• ./mvnw -f aggregator/pom.xml verify -pl :graph-compose-qa -am — full compile + javadoc gate
  • the qa visual/parity byte-identity suite (643): green. The layer-stack render (z-order,
    per-layer offsets, fill/overlay) is exercised by the qa LayerStack parity / demo suites.

Honest scope: a relocation with a host callback, not a decoupling — the collaborator is not
independently unit-testable (its per-layer step delegates back into the full compiler), so it
stays qa-covered like the pagination-cursor helpers. CompileContext is the first step toward
shrinking the remaining fat compile* signatures; retrofitting SplittableLeafCompiler to it
is a follow-up.