[Misc] Document data migrations and the document-queueing pattern

[manuelleduc]

Adds one OKF topic the map was missing: how a data migration is shaped, and the three rules that
otherwise make a document-queueing migration a silent no-op. Learned while validating the
XWIKI-20699 migration with real 17.10.9 → 18.7.0-SNAPSHOT upgrades on the supported database engines.

okf/architecture/data-migrations.md

• The queueing pattern — AbstractDocumentsMigration only writes rows into the async indexing
  queue and a separate TaskConsumer performs the change later, so the migration reporting success
  does not mean the data changed and the two halves fail independently.
• Pass "", never null, as the locale: LocaleUtils.toLocale(null) returns null rather than
  throwing, so every document resolves to Optional.empty() and nothing is queued.
• A TaskConsumer comparing resolved references needs @Named("current"), or the default resolver
  falls back to the default wiki and every subwiki is silently skipped — invisible on the main wiki.
• Guard possibly-empty strings in HQL with length(x) > 0, not <> '', because Oracle stores the
  empty string as NULL.
• A closing paragraph on validation: unit tests are structurally blind to the first two rules (a
  mocked resolver behaves the same whichever implementation is injected), so it takes a real
  FROM → TO upgrade, run on a subwiki as well as the main wiki, with the task queue drained before
  asserting, and on Oracle.

Scope

Deliberately kept to one file. The engine versions are marked volatile with a verify: recipe
pointing at the Release Notes and the Database Support Strategy page; only the policy is inline.
Harness-specific traps (building the distribution, readiness probing, extension.repositories,
patching the shipped Hibernate config) are left out — they are procedure, already covered by the
xwiki-build skill, or generic DBA trivia that would cost tokens in every session that loads this
file.

Grounding

The architecture claims are cited to the platform sources they come from and were re-read against
master while writing. The engine-specific behaviour was observed on a running instance.

Index and the injected map updated; plugin version bumped 1.1.7 → 1.1.8 across the manifests;
node scripts/validate.mjs and claude plugin validate ./xwiki pass.

[vmassol]

@manuelleduc I've updated it a bit to conform to best practices in the OKF. Please let me know what you want to do about this PR. Thx!