fix: keep control characters in content from acting as markers

[hownowstephen]

Conversion marks blockquotes, list indents, horizontal rules and <pre> blocks with the control characters \x00–\x06. The comment on those constants said the parser never emits control characters, but it does: raw ones pass through, and so do entities such as . Content containing them was read back as markup.

<p>see &#x01;quote&#x02; here</p>          before: "> see quote here"
                                           after:  "see quote here"

<p>x&#6;y</p><pre>code</pre>              before: "xcodey\n\n\x06"
                                           after:  "xy\n\ncode"

Text and attribute values are now read through withoutMarkers: text nodes, span text, <pre> text and getAttr. It does a byte scan first and only rebuilds a string when a marker is present. Markers are ASCII, so the scan cannot match inside a multi-byte rune.

I first tried a separate walk over the tree before conversion, but that cost about 3% and strings.Map on every string cost 8%. Reading through withoutMarkers costs +1.6% geomean when run interleaved against main, and only pathological is outside noise at +2.3%. Allocations are unchanged.

Output is byte-for-byte identical to main across the corpus under every link style.