feat(file): let the file name offer a type no content probe can find

[andiwand]

🤖 Generated with Claude Code

Closes #760.

The gap

FileType::markdown was the one document format detection could never reach.
detect_by_content is false for it — correctly, since CommonMark has no
signature and every plain text file is valid markdown — so list_file_types
never offered it and a .md came back as text_file. Worse, a .md that
happened to parse as CSV was offered as CSV.

Meanwhile the name was already at the boundary and thrown away:
abstract::File::disk_path() has carried it all along.

The change

Option (2) from the issue, which turned out to need no new plumbing.
open_strategy::file_type_by_name reads the extension off disk_path() and
returns the type only when the table says no content probe can produce it
(detect_by_content == false). Both detection paths consult it once the bytes
have decoded as text:

• list_file_types appends it last, so the preference path's reverse tries it
  first — a .md now lists [text_file, …, markdown].
• open_file tries it ahead of the csv/json/xml probes, so a markdown file
  that parses as CSV stays markdown.

A name never claims bytes: a .md holding a zip is still a zip, and
File::from_memory has no name, so it still needs FileType::markdown.

Verified

New odr.types_md and odr.a_misnamed_file_is_what_its_bytes_are.
FileTypeCapabilities.declaration_matches_the_engines held the old invariant —
"whatever detection sees, the table has to admit to" — and now admits the
name-offered type too; that test failing was the observable change the issue
predicted.

Full suite 1420 passed / 6 pre-existing skips. Reference output byte-identical
(HtmlOutputTests pins as_file_type, so it never went through detection).
Changed TU re-checked under -Werror and clang-tidy.