Treat u, s, and bdi as phrasing content (fixes #997)

[cpruijsen]

Fixes #997.

PHRASING_ELEMS already included B and BDO but not U, S, or BDI. When Readability wraps a DIV's phrasing children into <p> tags, those tags were treated as breaks, so

<div>Seigneur, mon Die<u>u</u> et mon salut.</div>

became

<p>Seigneur, mon Die</p><u>u</u><p> et mon salut.</p>

and excerpt stopped at "Seigneur, mon Die". Same split for <s>, <bdi>, and for <a> that contains <u>.

The sample in the issue is a <p>. Wrapping of consecutive phrasing children only runs on DIV; a <p> that already contains <u> is left alone. The incorrect output in the issue is the DIV-wrapping result, which is what the test page uses.

Summary

• Add U, S, and BDI to PHRASING_ELEMS (same list as B / BDO, matching HTML phrasing content).
• In _replaceBrs, do not pull a phrasing sibling that contains a <br><br> chain into the new <p>. Those chains become paragraphs next; wrapping them nests <p> inside <p>/<u>, and jsdom and JSDOMParser then disagree, so hukumusume cannot stay green on both parsers. A single nested <br> is still pulled in (it does not become a paragraph).
• Test page phrasing-u-s-bdi with the reporter's underline example plus <s> and <bdi> inside DIVs.

Decision

• Chose: only U, S, BDI. Alternative: also PICTURE / MAP / AREA / SLOT / TEMPLATE. Reasoning: those three are what PHRASING_ELEMS missing <u>, <s>, and <bdi> elements causes paragraph splitting #997 asked for, and the list already deliberately omits some phrasing tags (CANVAS, IFRAME, SVG, VIDEO). Happy to add more if you want.
• Chose: _replaceBrs skip when a phrasing node contains a <br><br> chain, rather than regenerating hukumusume/expected.html. Alternative: regenerate that fixture. Reasoning: JSDOMParser and jsdom then disagree on nested <p> inside <u>. Happy to switch if you want a parser picked as source of truth.

Test plan

• npm test -- -g phrasing-u-s-bdi (jsdom + JSDOMParser)
• npm test (existing pages, including hukumusume)
• npm run lint
• Optional: a live article that uses <u> / <s> / <bdi> inside DIV-wrapped paragraphs in Firefox Reader View