perf(sdk): benchmark and reduce the per-call cost of the paint walk #3078

Deferred from #3070 (thread), where the fix was judged to want a benchmark rather than eyeballing.

compositionPaintsAt is an uncached O(all stamped elements) walk, and the documented host recipe re-runs it on pointer movement and on every frame (animated artwork moves under a stationary cursor). Per call, per contained candidate:

• querySelectorAll + getBoundingClientRect on every stamped element — ensureHfIds stamps every body descendant, so addressableOnly barely shrinks it
• up to three getComputedStyle resolutions (isRenderedVisible, isOpacityVisible's inclusive re-read, styleInk)
• an eager full ancestor opacity walk plus a second depthOf ancestor walk
• a fresh root query per painted hit when fullBleedFraction > 0

On a ~1000-element composition that is order 100k rect/style reads per second, forcing layout flushes interleaved with GSAP's writes — janking the animation the overlay sits on.

Semantics-preserving wins identified in review:

• read computed style once per candidate and thread it through
• defer the visibility check into the lazy find predicate beside the ink test
• single-pass min-tracking instead of a full sort
• cache frame area per document

Why this needs a benchmark, not a patch. Two of those interact with the lazy-sampling property pinned by samples alpha lazily — the smallest painting box wins before the rest are read: deferring visibility changes how many candidates get style-resolved, and replacing the sort with min-tracking changes the order ink is evaluated in. Both are safe on paper and both could quietly cost more alpha samples than they save. Measure on a ~1000-element composition before and after.

The host-side half already landed in #3070: the guide now says to coalesce pointer / Alt / playhead into one gated rAF query and to short-circuit before the walk when the pointer is outside the composition's box.