perf: collapse per-child loop in findWithMatcher

[jvoisin]

findWithMatcher iterated over each element child of every input node and called m.MatchAll(c) on each, then append-spread-merged the returned slices into one result. For a node with K element children this was K independent cascadia tree walks, each starting with a fresh nil result slice.

This commit calls m.MatchAll(n) once on the input node itself. cascadia walks n plus all descendants in document order, so if n matches it is always the first element of the returned slice, strip it to preserve jQuery Find's "descendants only" semantics.

Benchmarks (-count=20, benchtime=500ms, linux/arm64):

                      sec/op           B/op             allocs/op
Find                  ~ (-10%)         -23.53%          -7.69%
FindWithinSelection   ~                -11.06%          -43.53%
FindNodes             -12.36%          ~                ~
Has                   ~                -17.20%          -7.14%
Add                   ~                 -5.13%          -9.09%
Closest               -37.34%          ~                ~
ClosestSelection      -43.77%          ~                ~
ClosestNodes          -36.56%          ~                ~
MetalReviewExample    ~                 -5.42%         -11.19%