Add automated conformance testing against html5lib-tests

[casoon]

Problem

Test coverage today is hand-written unit tests (in src/tree_builder.rs, src/tokenizer.rs) and end-to-end tests (src/lib.rs) targeting specific spec sections and scenarios ported from html-conform's own test matrix. That's solid for verifying individual algorithms, but there's no automated check against the html5lib-tests / WHATWG shared conformance corpus that every serious HTML5 parser (html5ever, html.parser, Validator.nu, browsers) is measured against. For a from-scratch, spec-transcribed parser, this is the single most valuable remaining quality signal — it would catch transcription mistakes and edge cases no amount of hand-written tests will reliably find.

Solution path

1. Vendor (or fetch via an xtask script, matching the existing xtask/gen-entities.py pattern) the tree-construction/*.dat files from https://github.com/html5lib/html5lib-tests — a well-known, widely-used format:
   • #data — input HTML.
   • #errors — expected parse errors (not applicable here, this crate has no diagnostics; skip this section).
   • #document — expected resulting tree, as an indented text dump.
   • Some files also have #document-fragment (fragment-parsing cases) and #script-off/#script-on — since this crate has no fragment parsing and always treats scripting as disabled, only the plain #document full-document cases apply. Document which sections/files are skipped and why (matching this project's existing "no silent exclusions" practice).
2. Write a small parser for the .dat format (it's simple — blank-line-separated sections) and a serializer that converts this crate's Document tree into the same indentation-based dump format html5lib-tests uses, so expected vs. actual can be compared as plain strings.
3. Add this as an integration test (e.g. tests/html5lib_conformance.rs) that runs each case through parse() and diffs the dump. Start with the simpler files (tests1.dat, tests2.dat) before the harder ones (tables01.dat, adoption01.dat; skip foreign-fragment.dat, it's fragment-only).
4. Expect (and explicitly track — e.g. a # known-failing allowlist, or #[ignore] with a documented reason) a nonzero number of failures at first, from the frameset/template gaps (see README overstates tree-construction completeness (frameset modes, <template> simplification undocumented) html5-parser#1) and any real transcription bugs the corpus surfaces. Fix what's fixable, document what's out of scope.
5. Wire into CI (.github/workflows/ci.yml) once the pass rate is stable, so future regressions get caught automatically.

This is a substantial task — plan to work through it in stages (vendor + harness first, then work through failures file by file), not as a single PR.