perf: replace recursive closure in Text() with iterative subtree walk

[jvoisin]

Text() built a self-referencing closure var f func(*html.Node); f = ... to recurse into descendants and accumulate text. The closure capture of its own variable forces the compiler to box f, and each recursive call pays a Go stack-frame setup.

This commit replaces it with an iterative depth-first walk using FirstChild/NextSibling/Parent pointers, anchored at each root in s.Nodes so the walk never escapes the original subtree.

Allocations are unchanged (the remaining 6/15/19 allocs come from strings.Builder grow-doubling), but ns/op drops by roughly 30-40% across shallow and deep subtrees:

name old time/op new time/op delta
Text-8 1.86µs 1.11µs -40.4% (h2, n=14)
TextDeep-8 126µs 86µs -31.6% (whole body)
TextManyShallow-8 40.9µs 29.1µs -28.9% (all anchors)

All p=0.000 over 10 runs on arm64. B/op and allocs/op are bit-for-bit identical in all three benchmarks.

[mna]

I'm not sold on that one, the resulting implementation is harder to read and has subtle conditions. I think just getting rid of the closure could be enough, and keep the recursive approach.

[jvoisin]

I guess it's a matter of taste, I prefer the explicit iterative version with comments.

But alright, I'll try a version using an outlined function instead of a closure, benchmark it, and will send a pull-request if it's improving things.

[jvoisin]

See #561