When an inline-relative is split by block-level descendants, the rect
computation only looked at one anonymous wrapper and returned empty
for everything else, sending abspos placement back to the initial
containing block. On Reddit this let an inline-relative ad host's <a>
overlay the entire viewport and steal clicks from the post gallery's
pager buttons.
Walk every descendant of the inline's real block container instead,
collecting fragments from any InlineNode part of the inline plus the
border-box rects of its in-flow Box descendants. Matches what other
engines expose via getClientRects() for split inlines.
While here, narrow the inline-CB detection to the triggers that
actually apply on non-atomic inlines: `position`, `filter`,
`backdrop-filter` (and their will-change hints). transform, contain
and the rest don't apply per their specs - the broader check would
have started routing the WPT contain-paint/contain-layout ib-split
tests through the inline path once the rect computation began
returning non-empty results.