Consolidate the mkdocs config reading, the docs root, and the stamp constants

[mairas]

From the review on PR #1: #1 (comment)

Three reviewers independently reached the same conclusion about the shape of the extraction.

• _Loader and its multi-constructor now exist verbatim in two modules, and four of six modules read mkdocs.yml — three of them by importing configured_languages from translation_status, which makes the translation report a library for the anchor checker. check-anchors is the one command docs.halos.fi runs and the one with nothing to do with translations. Consolidating is right, but not into translation_status. A mkdocs_config.py holding _Loader, configured_languages() and configured_base() is one small module with five callers.
• The docs root is treated five ways across six commands: a module constant in two, an inline Path("docs") in one, an inline Path("docs", language) in another, and a --docs option in the fifth. A reader cannot tell from any one module whether it is configurable, and the next command has five precedents to choose from.
• STAMP_KEY and blob_hash are declared independently in the module that writes the stamp and the one that reads it. STAMP_KEY is the on-disk format contract between them; PR Package the documentation checkers and add a blocking translation gate #1 had to edit blob_hash in both places to add --no-filters, which is exactly the failure this invites.
• The five entry states are bare strings with the set written out twice in the two renderers and "current" compared as a literal in four places. This is the vocabulary of a blocking gate: a typo in one comparison is a permanent pass or a permanent fail, and nothing type-checks it.
• chain_rule = () if allowed else HYPHEN_CHAINS.finditer(text) binds one name to two unrelated types, eight lines above an adjacent rule that uses a plain conditional.

[mairas]

One more config value belongs in this consolidation: site_dir.

check-anchors takes the built site as a positional argument, and the shared
workflow (halos-org/shared-workflows#39)
hard-codes site because that is the mkdocs default. A consuming repo that sets
site_dir in mkdocs.yml gets exit 2 on every run.

configured_base() already reads site_url from mkdocs.yml, so reading
site_dir from the same place would make the positional argument default to
whatever the build actually produces. Then neither the workflow nor a caller has
to know.

Measured while reviewing that PR: on exit 2 both messages go to stderr, so a
workflow piping stdout alone gets an empty diagnostic. That side is fixed in the
workflow with 2>&1, but the reason it mattered was this hard-coded path.