perf: skip dedup map in FindNodes

[jvoisin]

FindNodes went through mapNodes, which builds a map[*html.Node]bool to deduplicate results merged across source nodes, and wrapped every match in a one-element []*html.Node slice. Both are wasteful here: each source node maps to at most itself (when it is a descendant of the current selection), so distinct source nodes can never collide. Append the matching nodes directly into the result instead.

Duplicate input nodes still have to be discarded so the resulting Selection stays a proper set, which mapNodes previously handled via appendWithoutDuplicates. A cheap isInSlice check against the (small) result slice preserves that, and is placed before the more expensive sliceContains tree walk so duplicate inputs short-circuit early.

benchstat -count=12:

           |   old    |             new
           |  sec/op  |   sec/op     vs base

FindNodes-8 | 232.6µ | 131.0µ -43.68% (p=0.000 n=12)

           |   old    |             new
           |   B/op   |      B/op      vs base

FindNodes-8 | 12.023Ki | 2.164Ki -82.00% (p=0.000 n=12)

           |   old    |            new
           | allocs/op| allocs/op   vs base

FindNodes-8 | 85.000 | 9.000 -89.41% (p=0.000 n=12)