docs: restore the missing accents in the Spanish pages

[marevol]

Stacked on #509, because that PR also touches es/.../ds-database.rst. This PR's diff is against it.

Summary

Seven Spanish pages had lost most of their accents. The damage is partial rather
than stylistic - the same page reads Búsqueda in its title and Descripcion
two lines later, índice in one paragraph and indice in the next. It is a
defect, not a house style.

Share of non-blank lines carrying an accented character (áéíóúüñ, upper case,
¿¡):

Page	Before	After
config/datastore/ds-git.rst	0.005	0.195
config/datastore/ds-database.rst	0.017	0.271
config/llm-openai.rst	0.021	0.310
config/llm-gemini.rst	0.022	0.290
config/rag-chat.rst	0.040	0.386
config/datastore/ds-slack.rst	0.042	0.260
config/rate-limiting.rst	0.052	0.378

Measured the same way, the other 140 pages of es/15.9 run from 0.057 to 0.694
with a median of 0.296, and the only one below 0.09 is admin/index.rst, which
is almost entirely a toctree. These seven were the outliers.

(ds-database.rst starts at 0.017 rather than near zero only because the text
#509 adds to it is already accented.)

How the spellings were chosen

Not by guessing. Every accented word in the healthy Spanish pages of this
repository was collected and keyed by its unaccented form. A mapping was kept
only where exactly one accented form existed and the unaccented form is not
itself a Spanish word - 597 substitutions.

Two orthographic rules were applied on top: no Spanish word ends in an unaccented
-cion or -sion, and the plurals -ciones / -siones correctly carry none.

The remaining rare inflected forms and homographs were resolved one at a time
against their context: future and conditional (estaran → estarán,
eliminarian → eliminarían), preterite (elimino → eliminó, inicio →
inició where it is the verb and not the noun), imperative with an enclitic
pronoun (coloquelo → colóquelo, Administrela → Adminístrela), and
adjectives (explicito → explícito, matematicas → matemáticas).

Homographs left alone

• que, como, cuando, donde, cual stay unaccented as relatives and
  conjunctions. Only genuine interrogatives were accented: identifica qué paso ha fallado, and the heading Cómo funciona.
• esta / este stay unaccented as demonstratives; only verb forms became
  está / esté.
• si stays unaccented as the conditional.
• solo stays unaccented, per current RAE.
• envio was split by meaning: envió (verb) at one site, envío (noun) at
  another.

One bulk substitution was wrong and was reverted. Mejora continua
translates "Continuous Improvement", so the adjective is correct there and the
verb continúa is not.

Scope

Prose only, and 15.9 only.

• 1,079 protected fragments - every literal block body, inline `` span,
  URL and |Fess| substitution - were extracted from both revisions and compared:
  all byte-identical.
• The diff is 782 insertions against 782 deletions, so no line was added,
  removed or reflowed.
• Section title underlines are unaffected: an accent does not change a character
  count. tools/check_headings.py passes over the versions versions.json calls
  current, as do the tools unit tests and update_eol.py --check.

Spanish text inside code blocks was deliberately left as it is - for example the
curl payload in rag-chat.rst still reads explicame como instalar Fess, and
the placeholder `datastore://<id de la configuracion...>` in
ds-database.rst keeps its spelling because it sits inside an inline literal.

Not addressed

• ds-git.rst uses reindexiza, which is not standard Spanish (se reindexa
  would be). Not an accent problem, so out of scope here.
• llm-openai.rst and llm-gemini.rst use commita as a verb.
• Four pre-existing "title underline too short" warnings (rag-chat, rate-limiting)
  are present identically before and after this change.