docs: stop six English pages competing with themselves in search

[marevol]

Twelve English pages form six pairs that share a title, and every one of
them points at itself as canonical — so each pair competes for a single
search result and Google picks whichever it likes. These six pairs are the
only duplicate titles left across all 1,730 pages in the sitemap.

Five pairs are the same subject written twice

The shorter page exists only in English; the fuller one exists in all seven
languages.

English-only	counterpart (7 languages)	shared title	sizes
config/analyzer	config/admin-analyzer	Analyzer Configuration	14 vs 77 lines
config/role-setting	config/security-role	Role-Based Search Configuration	51 vs 72
config/virtual-host	config/security-virtual-host	Virtual Host	50 vs 75
config/search-config	config/search-advanced	Search-Related Settings	177 vs 393
config/index-backup	config/admin-index-backup	Index Management	13 vs 372

Each shorter page now names the fuller one as its canonical:

:canonical: admin-analyzer

The URLs stay and the content stays — an existing link keeps working.
What changes is that the pair is consolidated onto the page that is
translated everywhere, instead of the choice being made for us.

The value is a document name resolved against the declaring page's own
directory, which is why the identical line works in 15.7 and in 15.8.
Cutting the next version copies the directory verbatim; a value carrying
15.7 would leave the copy pointing at the previous version, and nothing
would say so.

The sixth pair is not a duplicate

/api/api-suggest and /api/admin/api-admin-suggest are different APIs —
one returns suggest words to a searcher, the other manages them and
requires an admin access token — that happen to share the English title
Suggest API. Japanese already tells them apart (サジェストAPI /
Suggest API). The admin one is retitled Admin Suggest API, using
the wording already on the page beside it (Admin API Overview, "the Admin
API permission"). Its rules are resized to the new title.

Scale, stated plainly

These twelve pages drew 276 impressions and 4 clicks over three months —
0.15% of the site's 187,000 impressions. This is tidying, not a traffic
measure. The cannibalisation is real (search-config 53 impressions +
search-advanced 39 = 92 impressions and zero clicks) but small.

Verification

• python3 tools/check_headings.py en/15.7 en/15.8 → exit 0
• python3 -m unittest discover -s tools -p 'test_*.py' → 23 tests, OK
• Built the affected pages with the site's Sphinx version and confirmed the
  emitted <link rel="canonical"> points where intended, that
  admin-analyzer and the two suggest pages still point at themselves, and
  that titles, og:url and the generated meta descriptions are unchanged.
• The :canonical: field is file-wide metadata: Sphinx removes it from the
  document, so it does not appear in any rendered page.

Reading the field is a change to the site generator; until that ships these
lines are inert metadata and the pages render exactly as they do today.