feat(file): name html as a file type, without a renderer for it

[andiwand]

🤖 Generated with Claude Code

Closes #763.

There is no FileType for html today, so an .html decodes as xml when it
happens to be well formed and as text when it is not — and a host that trusts
the answer shows the page's source instead of the page. iOS routes around it by
matching UTType against .html and .xhtml; Android has no such route and
opens an .html as an xml source tree.

This adds the row, and nothing else.

extensions	html, htm, xhtml
MIME types	text/html, application/xhtml+xml
category	text
capabilities	none

capabilities_by_file_type(hypertext_markup_language) declares no open and no
translate_html, so a host reads it as "the core names this and will not draw
it" and hands the file to the web view it already has — the shape psd, jp2,
wmf and emf took in 6.11.0. A passthrough renderer that emitted the file as
its own output would be of no use to such a host, and it would put the library
in the business of rewriting a page's document-relative links and assets.

Detection is by name: detect_by_content is false, as it is for markdown, since
html has no dependable signature and a probe for one is a probe for prose. What
matters is that file_type_by_file_extension("html") answers something other
than unknown. Detection from bytes is unchanged — a well formed .html still
lists as [text_file, xml].

FileType::hypertext_markup_language is appended to the enum, so the bindings
that mirror it by ordinal (jni, apple, python) gain the constant in the same
place; wasm derives its enum from the table already.

Test

FileTypeTable.html_is_named_but_not_decoded pins the aliases, the category,
the empty capabilities, and that opening a page as this type throws. The
table-wide invariants (one row per type, unique aliases, capabilities that do
not exceed the engines) cover the new row automatically.

Full odr_test suite: 1253 passed, 6 skipped (the pre-existing svm / encrypted
doc / wpd skips). Reference output is unchanged.