fix: encode carriage returns so HTML helpers round-trip them

[linkdata]

Fixes #202

Problem

AppendAttrValue delegates to html.EscapeString, which does not escape a carriage return, so a multiline attribute value read back from the DOM has its \r turned into \n (browser input-stream preprocessing rewrites raw CR/CRLF to LF before tokenization). Encoding the CR as 
 makes getAttribute report it unchanged.

Fix

AppendAttrValue now encodes each \r as the numeric character reference 
 (via a small shared appendEscapeCR helper). The parser decodes it back to CR after preprocessing, so the value round-trips through getAttribute. Attr, AppendAttr, WriteHTMLTag, and WriteHTMLInput inherit the fix.

Why WriteHTMLInner (textarea/pre) is not changed

The issue also asked to encode CR in textarea/pre inner content, but that turns out to be wrong for both:

• textarea — HTMLTextAreaElement.value normalizes every CR and CRLF to LF (WHATWG textarea value normalization), and jaws.js sends elem.value back to the server (jaws.js#L157). So a carriage return in textarea content can never round-trip through the value the browser reports, regardless of 
 encoding — <textarea>\n&#13;hello</textarea> gives textContent == "\rhello" but value == "\nhello".
• pre — pre permits phrasing content including <script> and comments, where character references are not decoded (script-data / comment parsing). Rewriting every CR would corrupt trusted markup, e.g. //x\rdoThing() would become the literal //x
doThing(), commenting out doThing().

So WriteHTMLInner writes innerHTML verbatim (unchanged behavior); the doc comment now explains the textarea/pre CR semantics and points callers who need CR retention to AppendAttrValue.

Tests

• Exact-output assertions for the new 
 attribute encoding (single CR and CRLF).
• TestAppendAttrValue_DOMRoundTrip — models getAttribute (input-stream preprocessing + character-reference decoding via stdlib html.UnescapeString) and confirms the logical value is recovered.
• TestWriteHTMLInner_TextareaValueNormalization — models HTMLTextAreaElement.value and pins the true contract: textarea content round-trips with CR/CRLF collapsed to LF.
• Newline-sensitive table now asserts textarea/pre content is written verbatim, including a pre+<script> regression case guarding against CR encoding inside nested markup.

Dependency surface stays stdlib-only. Package coverage remains 100% under -race; go vet, gofumpt, staticcheck, golangci-lint, and gosec are clean.