perf: presize dedup map in mapNodes

[jvoisin]

mapNodes is the shared engine behind most traversal helpers (Parents, Children, Next/Prev, Siblings, Find*, Contents, ...). It allocates a map[*html.Node]bool used by appendWithoutDuplicates to skip duplicates across per-node results. The map was created with no size hint, so it incurred grow-and-double rehashing as it filled up.

Pass len(nodes) as the size hint. That's a reasonable upper bound for the common case where each input yields 0 or 1 results (Parent, Next, Prev, Children, ...) and stays close enough for the multi-result cases (Parents, NextAll, PrevAll, ...) to still avoid most rehashes.

Benchmarks (count=10, benchtime=500ms, linux/arm64):

Notable per-bench wins (all p<0.01):
Children -32.35% sec/op
Prev -36.73% sec/op
Next -29.76% sec/op
NextFiltered -29.57% sec/op
PrevFiltered -27.12% sec/op
ParentsFilteredUntil* -25% sec/op, -21% B/op
FindWithinSelection -17.97% sec/op, -21% B/op
ContentsFiltered -20.34% sec/op
NextUntil -16.25% sec/op
ParentsUntil -15.42% sec/op
Contents -12.09% sec/op
Find -11.52% sec/op

FindSelection / FindNodes feed mapNodes a 448-element input, so the presize bumps B/op by ~5 KiB on those two benchmarks — but allocs/op drops from 91 to 85, because the single sized allocation replaces the grow-and-double sequence the map went through before.

No benchmark regressions were found.

[jvoisin]

Superseded by #563